Fi|dZddlmZddlmZmZddlmZddlm Z ddl Z ddl Z ddl Z ddlmcmZddlmZddlmZdd lmZdd lmZmZmZdd lmZmZdd lm Z m!Z!d dl"m#Z#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z)m*Z*m+Z+d dl,m-Z-GddeZ.Gdde.Z/Gdde/Z0Gdde/Z1Gdde/Z2Gdde2Z3Gdde0e2Z4Gdde3Z5y) a Generate predictions using the Segment Anything Model (SAM). SAM is an advanced image segmentation model offering features like promptable segmentation and zero-shot performance. This module contains the implementation of the prediction logic and auxiliary utilities required to perform segmentation using SAM. It forms an integral part of the Ultralytics framework and is designed for high-performance, real-time image segmentation tasks. ) annotations) OrderedDict defaultdict)deepcopy)AnyN) LetterBox) BasePredictor)Results) DEFAULT_CFGLOGGERops)box_ioumask_iou) select_devicesmart_inference_mode) batch_iteratorbatched_mask_to_boxbuild_all_layer_point_gridscalculate_stability_scoregenerate_crop_boxesis_box_near_crop_edgeremove_small_regionsuncrop_boxes_xyxy uncrop_masks)PromptceZdZdZdZeddfdfd ZdZdZddZ ddZ dd Z dd Z dd Z dd Zd ZdZdZfdZdZdZdZeddZe ddZxZS) PredictoraPredictor class for SAM, enabling real-time image segmentation with promptable capabilities. This class extends BasePredictor and implements the Segment Anything Model (SAM) for advanced image segmentation tasks. It supports various input prompts like points, bounding boxes, and masks for fine-grained control over segmentation results. Attributes: args (SimpleNamespace): Configuration arguments for the predictor. model (torch.nn.Module): The loaded SAM model. device (torch.device): The device (CPU or GPU) on which the model is loaded. im (torch.Tensor): The preprocessed input image. features (torch.Tensor): Extracted image features. prompts (dict[str, Any]): Dictionary to store various types of prompts (e.g., bboxes, points, masks). segment_all (bool): Flag to indicate if full image segmentation should be performed. mean (torch.Tensor): Mean values for image normalization. std (torch.Tensor): Standard deviation values for image normalization. Methods: preprocess: Prepare input images for model inference. pre_transform: Perform initial transformations on the input image. inference: Perform segmentation inference based on input prompts. prompt_inference: Internal function for prompt-based segmentation inference. generate: Generate segmentation masks for an entire image. setup_model: Initialize the SAM model for inference. get_model: Build and return a SAM model. postprocess: Post-process model outputs to generate final results. setup_source: Set up the data source for inference. set_image: Set and preprocess a single image for inference. get_im_features: Extract image features using the SAM image encoder. set_prompts: Set prompts for subsequent inference. reset_image: Reset the current image and its features. remove_small_regions: Remove small disconnected regions and holes from masks. Examples: >>> predictor = Predictor() >>> predictor.setup_model(model_path="sam_model.pt") >>> predictor.set_image("image.jpg") >>> bboxes = [[100, 100, 200, 200]] >>> results = predictor(bboxes=bboxes) Nc|i}|jtdddt| |||d|j_d|_d|_i|_d|_ y)aeInitialize the Predictor with configuration, overrides, and callbacks. Sets up the Predictor object for SAM (Segment Anything Model) and applies any configuration overrides or callbacks provided. Initializes task-specific settings for SAM, such as retina_masks being set to True for optimal results. Args: cfg (dict): Configuration dictionary containing default settings. overrides (dict | None): Dictionary of values to override default configuration. _callbacks (dict | None): Dictionary of callback functions to customize behavior. Nsegmentpredictr)taskmodebatchTF) updatedictsuper__init__args retina_masksimfeaturesprompts segment_allselfcfg overrides _callbacks __class__s `/var/www/pod-logistic/pod-ai/venv/lib/python3.12/site-packages/ultralytics/models/sam/predict.pyr)zPredictor.__init__Wsa  I99AFG i4!%    c |j |jSt|tj }|rgt j |j |}|ddddfjd}t j|}tj|}|j|j}|r||jz |jz }|jjr|j!}|S|j#}|S)avPreprocess the input image for model inference. This method prepares the input image by applying transformations and normalization. It supports both torch.Tensor and list of np.ndarray as input formats. For OpenCV-loaded images, the input is typically BGR and is converted to RGB during preprocessing. Args: im (torch.Tensor | list[np.ndarray]): Input image(s) in BCHW tensor format or a list of HWC NumPy arrays. NumPy arrays are expected to be in BGR order (as returned by OpenCV) and will be converted to RGB. Returns: (torch.Tensor): The preprocessed image tensor, normalized and converted to the appropriate dtype. Examples: >>> predictor = Predictor() >>> image = torch.rand(1, 3, 640, 640) >>> preprocessed_image = predictor.preprocess(image) N.)rr)r, isinstancetorchTensornpstack pre_transform transposeascontiguousarray from_numpytodevicemeanstdmodelfp16halffloat)r1r, not_tensors r6 preprocesszPredictor.preprocessms& 77 77N#B 55 $,,R01BC2I((6B%%b)B!!"%B UU4;;  tyy.DHH,B**//RWWY 02xxz r7ct|dk(sJdt|jdd}|Dcgc] }|| c}Scc}w)0Perform initial transformations on the input image for preprocessing. This method applies transformations such as resizing to prepare the image for further preprocessing. Currently, batched inference is not supported; hence the list length should be 1. Args: im (list[np.ndarray]): List containing a single image in HWC numpy array format. Returns: (list[np.ndarray]): List containing the transformed image. Raises: AssertionError: If the input list contains more than one image. Examples: >>> predictor = Predictor() >>> image = np.random.rand(480, 640, 3) # Single HWC image >>> transformed = predictor.pre_transform([image]) >>> print(len(transformed)) 1 r6SAM model does not currently support batched inferenceF)autocenterimagelenrimgszr1r, letterboxxs r6rAzPredictor.pre_transformsE,2w!|UUU|djjuUC ,./q "///sAcd|jjd|}|jjd|}|jjd|}|jjd|}td|||fDr|j|g|i|S|j ||||||S)a=Perform image segmentation inference based on the given input cues, using the currently loaded image. This method leverages SAM's (Segment Anything Model) architecture consisting of image encoder, prompt encoder, and mask decoder for real-time and promptable segmentation tasks. Args: im (torch.Tensor): The preprocessed input image in tensor format, with shape (N, C, H, W). bboxes (np.ndarray | list | None): Bounding boxes with shape (N, 4), in XYXY format. points (np.ndarray | list | None): Points indicating object locations with shape (N, 2), in pixels. labels (np.ndarray | list | None): Labels for point prompts, shape (N,). 1 = foreground, 0 = background. masks (np.ndarray | None): Low-resolution masks from previous predictions, shape (N, H, W). For SAM H=W=256. multimask_output (bool): Flag to return multiple masks. Helpful for ambiguous prompts. *args (Any): Additional positional arguments. **kwargs (Any): Additional keyword arguments. Returns: pred_masks (torch.Tensor): The output masks in shape (C, H, W), where C is the number of generated masks. pred_scores (torch.Tensor): An array of length C containing quality scores predicted by the model for each mask. Examples: >>> predictor = Predictor() >>> predictor.setup_model(model_path="sam_model.pt") >>> predictor.set_image("image.jpg") >>> results = predictor(bboxes=[[0, 0, 100, 100]]) bboxespointsmaskslabelsc3$K|]}|du ywN.0is r6 z&Predictor.inference..s:QqDy:s)r.popallgenerateprompt_inference) r1r,r]r^r`r_multimask_outputr*kwargss r6 inferencezPredictor.inferences8!!(F3!!(F3   %0!!(F3 :665"9: : 4==5d5f5 5$$RHXYYr7c|j|j|n |j}|j|jdd|jddjdd||||}|j |g||S)aPerform image segmentation inference based on input cues using SAM's specialized architecture. This internal function leverages the Segment Anything Model (SAM) for prompt-based, real-time segmentation. It processes various input prompts such as bounding boxes, points, and masks to generate segmentation masks. Args: im (torch.Tensor): Preprocessed input image tensor with shape (N, C, H, W). bboxes (np.ndarray | list | None): Bounding boxes in XYXY format with shape (N, 4). points (np.ndarray | list | None): Points indicating object locations with shape (N, 2) or (N, num_points, 2), in pixels. labels (np.ndarray | list | None): Point prompt labels with shape (N) or (N, num_points). 1 for foreground, 0 for background. masks (np.ndarray | None): Low-res masks from previous predictions with shape (N, H, W). For SAM, H=W=256. multimask_output (bool): Flag to return multiple masks for ambiguous prompts. Returns: pred_masks (torch.Tensor): Output masks with shape (C, H, W), where C is the number of generated masks. pred_scores (torch.Tensor): Quality scores predicted by the model for each mask, with length C. Examples: >>> predictor = Predictor() >>> im = torch.rand(1, 3, 1024, 1024) >>> bboxes = [[100, 100, 200, 200]] >>> masks, scores, logits = predictor.prompt_inference(im, bboxes=bboxes) Nr;rr)r-get_im_features_prepare_promptsshaper%_inference_features) r1r,r]r^r`r_rlr-r.s r6rkzPredictor.prompt_inferences404}}/D4''+$--'' djjmA6F6L6LRa6PRXZ`bhjop't''M7M.>t{{SEvvu,,Ks-Ic (ddl} d|_|jdd\} }t| |f||\}}| t |||}ggggf\}}}}t ||D]\}}|\}}}}||z ||z }}t j||z|j}tj||gg}tj|d||||f| |fdd }|||z} ggg}#}"}!t|| D]\}$|j||$d \}%}&tj|%d||fdd d}%|&|kD}'|%|'|&|'}&}%t|%|j j"| }(|(| kD}'|%|'|&|'}&}%|%|j j"kD}%t%|%j'})t)|)|dd|| g}*t j*|*s|)|*|%|*|&|*}&}%})|!j-|%|#j-|)|"j-|&t j.|!}!t j.|#}#t j.|"}"| j0j3|#|"|j4j6}+t9|#|+|}#t;|!|+|| |}!|"|+}"|j-|!|j-|#|j-|"|j-|j=|!jdt j.|}t j.|}t j.|}t j.|}t?|d kDr1d |z },| j0j3||,| }+||+||+||+}}}|||fS) aPPerform image segmentation using the Segment Anything Model (SAM). This method segments an entire image into constituent parts by leveraging SAM's advanced architecture and real-time performance capabilities. It can optionally work on image crops for finer segmentation. Args: im (torch.Tensor): Input tensor representing the preprocessed image with shape (N, C, H, W). crop_n_layers (int): Number of layers for additional mask predictions on image crops. crop_overlap_ratio (float): Overlap between crops, scaled down in subsequent layers. crop_downscale_factor (int): Scaling factor for sampled points-per-side in each layer. point_grids (list[np.ndarray] | None): Custom grids for point sampling normalized to [0,1]. points_stride (int): Number of points to sample along each side of the image. points_batch_size (int): Batch size for the number of points processed simultaneously. conf_thres (float): Confidence threshold [0,1] for filtering based on mask quality prediction. stability_score_thresh (float): Stability threshold [0,1] for mask filtering based on stability. stability_score_offset (float): Offset value for calculating stability score. crop_nms_thresh (float): IoU cutoff for NMS to remove duplicate masks between crops. Returns: pred_masks (torch.Tensor): Segmented masks with shape (N, H, W). pred_scores (torch.Tensor): Confidence scores for each mask with shape (N,). pred_bboxes (torch.Tensor): Bounding boxes for each mask with shape (N, 4). Examples: >>> predictor = Predictor() >>> im = torch.rand(1, 3, 1024, 1024) # Example input image >>> masks, scores, boxes = predictor.generate(im) rNTr;rF.bilinearF)r$ align_corners)r^rlr) torchvisionr/rrrrzipr=rrFr?arrayF interpolaterrkrrImask_thresholdrrLrriappendcatr nmsr*iourrexpandrW)-r1r, crop_n_layerscrop_overlap_ratiocrop_downscale_factor point_grids points_stridepoints_batch_size conf_thresstability_score_threshstability_score_offsetcrop_nms_threshrihiw crop_regions layer_idxsrr pred_bboxes region_areas crop_region layer_idxx1y1x2y2wharea points_scalecrop_impoints_for_image crop_masks crop_scores crop_bboxesr^ pred_mask pred_scoreidxstability_score pred_bbox keep_maskkeepscoress- r6rjzPredictor.generateIsT !"B#6BxPb#c j  5m]TijK=?R^: Kl&), &C. B "K(NBB7BGqA<<Abii8D88aVH-LmmBsBrE2b5'8$9B8*dijG*95 D 35r2[ J+,=?OP / (,(=(=gfgk(=(l% :MM)D/Aq6 bghijk  :-(1# 3: ";tzz88:P#&(>>(1# 3: % (A(AA / :@@B 29kAqRTVX>ZZ yy+7@7KYW`Macmnwcx*yI!!),""9-"":.- /2:.J))K0K))K0K??&&{KOD+K,={KK%j&6 RLJ%d+K   j )   { +   { +    J,<,>> predictor = Predictor() >>> predictor.setup_model(model=sam_model, verbose=True) )verboseN)g33333^@gR]@gRY@r9r)g(\2M@g(\L@gL@sam T)rr*rF get_modelrErKrLevalrIr=rviewrGrHformatstriderJ done_warmupfloat16float32r)r1rIrrFs r6 setup_modelzPredictor.setup_modelstyy//A =NN$E  $  EKKM   LL!:;@@QJMMfU << 78==b!QGJJ6R"  )).. ,0JJOO5==r7cFddlm}||jjS)zPRetrieve or build the Segment Anything Model (SAM) for image segmentation tasks.r build_sambuildrr*rIr1rs r6rzPredictor.get_model$))r7c J|dd\}}|jr|dnd}ttdt|jdD}t |t stj|ddddf}g}t|g||jdD]\} } } | jddk(r%dtjd|j}} n$tj| dj| jddd d} | |j j"kD} |?tj$|jdd|j| jd }n t'| }tj(|jdtj*|j } ||j,j.kD} tj0||dddf| dddfgd | }| | } |j3t5| | || | d |_|S)aPost-process SAM's inference outputs to generate object detection masks and bounding boxes. This method scales masks and boxes to the original image size and applies a threshold to the mask predictions. It leverages SAM's advanced architecture for real-time, promptable segmentation tasks. Args: preds (tuple): The output from SAM model inference, containing: - pred_masks (torch.Tensor): Predicted masks with shape (N, 1, H, W). - pred_scores (torch.Tensor): Confidence scores for each mask with shape (N, 1). - pred_bboxes (torch.Tensor, optional): Predicted bounding boxes if segment_all is True. img (torch.Tensor): The processed input image tensor with shape (C, H, W). orig_imgs (list[np.ndarray] | torch.Tensor): The original, unprocessed images. Returns: (list[Results]): List of Results objects containing detection masks, bounding boxes, and other metadata for each processed image. Examples: >>> predictor = Predictor() >>> preds = predictor.inference(img) >>> results = predictor.postprocess(preds, img, orig_imgs) Nr;c32K|]}t|ywrbstrrds r6rgz(Predictor.postprocess..sJ!s1vJr.r9rrFpaddingrdimpathnamesr_rv)r/r' enumeraterangerrr<listr convert_torch2numpy_batchrr%r=zerosrF scale_masksrLrIr scale_boxesrarangerr*confrrr )r1predsimg orig_imgsrrrrresultsr_orig_imgimg_pathclsrs r6 postprocesszPredictor.postprocesss0#() K"&"2"2eAh YJuZ5E5Ea5H/IJJK)T*55i@ddKI),j\9djjQRm)T j %E8X{{1~"%)5;;vjFWFW+X{d (9(9(;X^^BQ=OY^_`ab 9 99*"%//#))AB-ARARATV^VdVdns"tK"5e">> predictor = Predictor() >>> predictor.set_image("path/to/image.jpg") >>> predictor.set_image(cv2.imread("path/to/image.jpg")) Notes: - This method should be called before performing inference on a new image. - The extracted features are stored in the `self.features` attribute for later use. Nrz,`set_image` only supports setting one image!)rIr setup_sourcerWdatasetrNrpr-)r1rUr%r,s r6 set_imagezPredictor.set_images. ::      % 4<< A%U'UU%\\ Eq*B 004DM  r7c2|yt|||jt|jt t fr|jd|jdk(sJd|jd|jj|jy)z)Set up the data source for SAM inference.Nrrz3SAM models only support square image size, but got r) r(r rr<rXtuplerrI set_imgsz)r1sourcer5s r6r zPredictor.setup_source3s} >  VT[[1$**udm4A$**UV-9W A$**Q O W TZZ(r7c8|jj|S)zZExtract image features using the SAM model's image encoder for subsequent mask prediction.)rI image_encoderr1r,s r6rpzPredictor.get_im_features=szz''++r7c||_y)z0Set prompts for subsequent inference operations.N)r.)r1r.s r6 set_promptszPredictor.set_promptsAs  r7c d|_d|_y)zQReset the current image and its features, clearing them for subsequent inference.N)r,r-r1s r6 reset_imagezPredictor.reset_imageEs r7cddl}|jddk(r|Sg}g}|D]}|jjj t j }t||d\}}| }t||d\}}|xr| }|jtj|jd|jt|tj|d}t|} |jj!| jtj||} || j#|j$|j&| fS)aWRemove small disconnected regions and holes from segmentation masks. This function performs post-processing on segmentation masks generated by the Segment Anything Model (SAM). It removes small disconnected regions and holes from the input masks, and then performs Non-Maximum Suppression (NMS) to eliminate any newly created duplicate boxes. Args: masks (torch.Tensor): Segmentation masks to be processed, with shape (N, H, W) where N is the number of masks, H is height, and W is width. min_area (int): Minimum area threshold for removing disconnected regions and holes. Regions smaller than this will be removed. nms_thresh (float): IoU threshold for the NMS algorithm to remove duplicate boxes. Returns: new_masks (torch.Tensor): Processed masks with small regions removed, shape (N, H, W). keep (list[int]): Indices of remaining masks after NMS, for filtering corresponding boxes. Examples: >>> masks = torch.rand(5, 640, 640) > 0.5 # 5 random binary masks >>> new_masks, keep = remove_small_regions(masks, min_area=100, nms_thresh=0.7) >>> print(f"Original masks: {masks.shape}, Processed masks: {new_masks.shape}") >>> print(f"Indices of kept masks: {keep}") rNholesr$islandsrrFr)rrrcpunumpyastyper?rrrr=r unsqueezerLrrr rrErFr) r_min_area nms_threshr new_masksrmaskchanged unchangedrvrs r6rzPredictor.remove_small_regionsJs.2  ;;q>Q L  ,D88:##%,,RXX6D0xgNMD'# I0xiPMD'!1'kI   U__T2<>> predictor = SAM2Predictor(cfg) >>> predictor.set_image("path/to/image.jpg") >>> bboxes = [[100, 100, 200, 200]] >>> result = predictor(bboxes=bboxes)[0] >>> print(f"Predicted {len(result.masks)} masks with average score {result.boxes.conf.mean():.2f}") ))r:)r;)r,r,rcFddlm}||jjS)zYRetrieve and initialize the Segment Anything Model 2 (SAM2) for image segmentation tasks.rrrrs r6rzSAM2Predictor.get_modelrr7c|t|||||||\}}}}||jddd}tjddggtj |j j|jdd}|3tj||gd}tj||gd}n||}}|||fS)aiPrepare and transform the input prompts for processing based on the destination shape. Args: dst_shape (tuple[int, int]): The target shape (height, width) for the prompts. src_shape (tuple[int, int]): The source shape (height, width) of the input image. bboxes (np.ndarray | list | None): Bounding boxes in XYXY format with shape (N, 4). points (np.ndarray | list | None): Points indicating object locations with shape (N, 2) or (N, num_points, 2), in pixels. labels (np.ndarray | list | None): Point prompt labels with shape (N,) or (N, num_points). 1 for foreground, 0 for background. masks (list | np.ndarray | None): Masks for the objects, where each mask is a 2D array. Returns: points (torch.Tensor | None): Transformed points. labels (torch.Tensor | None): Transformed labels. masks (torch.Tensor | None): Transformed masks. Raises: AssertionError: If the number of points don't match the number of labels, in case labels were passed. r9r;r:rrrr) r(rqrr=rrrFrrrr) r1rrr]r^r`r_ bbox_labelsr5s r6rqzSAM2Predictor._prepare_promptss*).(@IW]_egmot(u%  [[Q*F,,Axu{{6==Y``agamamnoaprtuK!FF#3;K#8a@!'vu$$r7c t||dDcgc]5}|jDcgc]}t||j|zz c}7c}}|_ycc}wcc}}w)z9Set up the data source and image size for SAM2 inference.g??rN)r(r rXintr_bb_feat_sizesr1rrfr[r5s r6r zSAM2Predictor.setup_sourcesK V$YjkTUDJJOqAq$9 :OkOksA"A AAc |jj|}|jj|\}}}}|jjr|d|jjz|d<t ||j Dcgc]*\}}|jdddjddg|,}}}|d|dddScc}}w)zLExtract image features from the SAM image encoder for subsequent processing.r9rr;rN) image_embedhigh_res_feats) rI forward_image_prepare_backbone_featuresdirectly_add_no_mem_embed no_mem_embedrrCpermuter)r1r, backbone_out_ vision_featsfeat feat_sizefeatss r6rpzSAM2Predictor.get_im_featuresszz//3 $ E El S<A :: / /+B/$**2I2IIL QTUacgcvcvQw >MdI &DLLAq ! & &q" 9y 9   %RyE#2JGG s /Cc |||fnd}|jj|d|\}}|duxr|djddkD} d} t|tr-|dD cgc]} | |j d} } |d|g}|jj ||jjj|||| | \} } }}| jdd| jddfScc} w)avPerform inference on image features using the SAM2 model. Args: features (torch.Tensor | dict[str, Any]): Extracted image features with shape (B, C, H, W) from the SAM2 model image encoder. Can also be a dict with 'image_embed' (torch.Tensor) of shape (B, C, H, W) and 'high_res_feats' (list[torch.Tensor]) high-resolution feature maps from the backbone. points (np.ndarray | list[list[float]] | None): Object location points with shape (N, 2), in pixels. labels (np.ndarray | list[int] | None): Point prompt labels with shape (N,). 1 = foreground, 0 = background. masks (list[np.ndarray] | np.ndarray | None): Masks for the objects, where each mask is a 2D array. multimask_output (bool): Flag to return multiple masks for ambiguous prompts. img_idx (int): Index of the image in the batch to process. Returns: pred_masks (torch.Tensor): Output masks with shape (C, H, W), where C is the number of generated masks. pred_scores (torch.Tensor): Quality scores for each mask, with length C. NrurrrGrF)rwrxryrzrl repeat_imagehigh_res_features) rIsam_prompt_encoderrrr<r'r sam_mask_decoderr}r~)r1r-r^r`r_rlimg_idxrr batched_moderU feat_levelrrrNs r6rsz!SAM2Predictor._inference_features s%2&,%7&&!T.2jj.K.K/L/ ++ T)DfQiooa.@1.D   h %T\]mTn ojG!4!>!>q!A o  o .y9H(, (C(C%ZZ22??A%6$4-%/)D) % KA!!!Q')<)>> predictor = SAM2VideoPredictor(cfg=DEFAULT_CFG) >>> predictor.set_image("path/to/video_frame.jpg") >>> bboxes = [[100, 100, 200, 200]] >>> results = predictor(bboxes=bboxes) Notes: The `fill_hole_area` attribute is defined but not used in the current implementation. Nct||||i|_d|_d|_d|_|j dj|jd|_ y)aKInitialize the predictor with configuration and optional overrides. This constructor initializes the SAM2VideoPredictor with a given configuration, applies any specified overrides, and sets up the inference state along with certain flags that control the behavior of the predictor. Args: cfg (dict): Configuration dictionary containing default settings. overrides (dict | None): Dictionary of values to override default configuration. _callbacks (dict | None): Dictionary of callback functions to customize behavior. TFon_predict_startN) r(r)inference_statenon_overlap_masksclear_non_cond_mem_around_input clear_non_cond_mem_for_multi_obj callbacksr init_stateclear_non_cond_memr0s r6r)zSAM2VideoPredictor.__init__`sZ i4!!%/4,05- )*11$//B"&r7cFt|}|jd|S)zRetrieve and configure the model with binarization enabled. Notes: This method overrides the base class implementation to set the binarize flag to True. T)r(r set_binarize)r1rIr5s r6rzSAM2VideoPredictor.get_modelss% !# 4  r7c |jjd|}|jjd|}|jjd|}|jj}||jd<|jd}t |ddk(r|j |jdd |jd djd d||||\}}}|7tt |D]}|j|||g||g| !n3|1tt |D]}|j|||g| |j|jd } t |jd} t |ddk(r td|| dvr9d} || |} |jri|js| d krX|j|nF|| dvr d} || |} n4d} |j!||| dd d dd} | || |<|j#||j%|| | |jdj'|| dj)dd } | | |j*j,kDj/ddkD} | t1j2| jd| j4| j6fS)abPerform image segmentation inference based on the given input cues, using the currently loaded image. This method leverages SAM's (Segment Anything Model) architecture consisting of image encoder, prompt encoder, and mask decoder for real-time and promptable segmentation tasks. Args: im (torch.Tensor): The preprocessed input image in tensor format, with shape (N, C, H, W). bboxes (np.ndarray | list, optional): Bounding boxes with shape (N, 4), in XYXY format. points (np.ndarray | list, optional): Points indicating object locations with shape (N, 2), in pixels. labels (np.ndarray | list, optional): Labels for point prompts, shape (N, ). 1 = foreground, 0 = background. masks (np.ndarray, optional): Low-resolution masks from previous predictions shape (N,H,W). For SAM H=W=256. Returns: pred_masks (torch.Tensor): The output masks in shape CxHxW, where C is the number of generated masks. pred_scores (torch.Tensor): An array of length C containing predicted quality scores for each mask. r]r^r_r, output_dictcond_frame_outputsrr;Nr)obj_idr^r` frame_idx)rkr_rlconsolidated_frame_inds obj_idx_to_id/No points are provided; please add points firstnon_cond_frame_outputsFT)rirl batch_sizeis_init_cond_frame point_inputs mask_inputsreverserun_mem_encoderframes_already_trackedrrr;r)r.rhr framer_rWrqrrr%radd_new_promptspropagate_in_video_preflight RuntimeErrorrarb _clear_non_cond_mem_around_input_run_single_frame_inference_prune_non_cond_memory_add_output_per_objectrr~rIrsumr=rrrF)r1r,r]r^r`r_ryrirfrmrq storage_key current_outrs r6rnzSAM2VideoPredictor.inference}s"!!(F3!!(F3   %0 ""%'T"**=9 {/0 1Q 6$($9$9 djjmA.44Ra8&&&RW% !FFE!s6{+lA((&!+fVWUXkej(kl"s5z*VA((su(UV ))+"&"6"67P"Q--o>? {/0 1Q 6PQ Q +,@A A.K%k259K339^9^blpqbq55e< -.FG G2K%k259K2K::'%#(!  $; K/:K $U +  ' ' . ##E; D 56==eD .66q!< djj.G.G!G L LV TWX XY 5::j&6&6q&9AQAQZdZkZklllr7c:t||||}|jr||D]w}|jt |jdk(r(|j j |jjjdd|j_y|S)a3Post-process the predictions to apply non-overlapping constraints if required. This method extends the post-processing functionality by applying non-overlapping constraints to the predicted masks if the `non_overlap_masks` flag is set to True. This ensures that the masks do not overlap, which can be useful for certain applications. Args: preds (tuple[torch.Tensor, torch.Tensor]): The predicted masks and scores from the model. img (torch.Tensor): The processed image tensor. orig_imgs (list[np.ndarray]): The original images before processing. Returns: (list): The post-processed predictions. Notes: If `non_overlap_masks` is True, the method applies constraints to ensure non-overlapping masks. r) r(rr`r_rWrI"_apply_non_overlapping_constraintsdatar )r1rrrrresultr5s r6rzSAM2VideoPredictor.postprocesss$'%eS)<  ! !! u<<'3v||+<+A$(JJ$Q$QRXR^R^RcRcRmRmnoRp$qrs$t ! ur7c |xs |j}|du|duz sJd|j||}d}d} |||d}||d||<d} ||d||<|| |j|d||dv} |d|} |d|} | xs|jj} | rd nd }d}|| |j |xs*| d j |xs| d j |}|Z|j d I|d j |j|jjd k( }|jdd|j| |d| ||dd|| }|| ||<|j|| d|}|d jdd}|jddtjd|j|jfS)avAdd new points or masks to a specific frame for a given object ID. This method updates the inference state with new prompts (points or masks) for a specified object and frame index. It ensures that the prompts are either points or masks, but not both, and updates the internal state accordingly. It also handles the generation of new segmentations based on the provided prompts and the existing state. Args: obj_id (int): The ID of the object to which the prompts are associated. points (torch.Tensor, optional): The coordinates of the points of interest. labels (torch.Tensor, optional): The labels corresponding to the points. masks (torch.Tensor, optional): Binary masks for the object. frame_idx (int, optional): The index of the frame to which the prompts are applied. inference_state (dict[str, Any], optional): The current inference state. If None, uses the instance's inference state. Returns: pred_masks (torch.Tensor): The flattened predicted masks. pred_scores (torch.Tensor): A tensor of ones indicating the number of objects. Raises: AssertionError: If both `masks` and `points` are provided, or neither is provided. Notes: - Only one type of prompt (either points or masks) can be added per call. - If the frame is being tracked for the first time, it is treated as an initial conditioning frame. - The method handles the consolidation of outputs and resizing of masks to the original video resolution. Nz@'masks' and 'points' prompts are not compatible with each other.point_inputs_per_obj point_coords point_labelsmask_inputs_per_objrwoutput_dict_per_objtemp_output_dict_per_objrjrprcuda)rF non_blockingg@g@@rF) rirlrqrrrsrtrurvprev_sam_mask_logitsr_is_condrvr_rr)r__obj_id_to_idxrhrI!add_all_frames_to_correct_as_condgetrErFtypeclamp_r~#_consolidate_temp_output_across_objr~r=rr)r1rkr^r`r_rlr_obj_idxrspop_keyrrobj_output_dictobj_temp_output_dictrrrprev_outrconsolidated_outrs r6rzz"SAM2VideoPredictor.add_new_promptsskL*AT-A-A &D.1u3uu1%%fo> (  ,2FKLJVO2 3G 'o>V.WW)*?@I./IJ7S%T (T(T.5*;S  $  #$[155i@L"#78<FTrjrprr_rNrrr)r_rWsetvaluesr&keysrrrarbr}clearrhdiscard)r1r_rqrrirmrrtemp_frame_indsrrlrrall_consolidated_frame_indsinput_frames_indspoint_inputs_per_framemask_inputs_per_frames r6r{z/SAM2VideoPredictor.propagate_in_video_preflightMs*AT-A-A26./9: $33M#N %m4 #22K"L$ :G29.?WK"eO(@(G(G(I Q$&&';K'H'M'M'OP Q #K 0 7 7 H, E #'#K#KwVe$L$ 7G K(3++I7Gfu+v77T=b=bfptufu99)D E)A(G(G(I :$$[1779 :/ :8%%9: GI 0 1 5 5i F G./DELLN OO,-AB O  89==iN O O11EF QI ,@ AA AA #$< = E Ei P Q $$8 9OO ,V 48>OO ,W 5)-oo.N)OOI &?AO2 3G <>@O1 27 ;&(*,?O1 27 ; ')*,DO6 7 @N+F84,,;I,F+GHEF r7c | xs |j} |j| d|\} } } ||J|jj||| | | |||| d|||  }|d}|F|j t j |j|jjdk(|d<|j|d| |d<|S)aRun tracking on a single frame based on current inputs and previous memory. Args: output_dict (dict): The dictionary containing the output states of the tracking process. frame_idx (int): The index of the current frame. batch_size (int): The batch size for processing the frame. is_init_cond_frame (bool): Indicates if the current frame is an initial conditioning frame. point_inputs (dict | None): Input points and their labels. mask_inputs (torch.Tensor | None): Input binary masks. reverse (bool): Indicates if the tracking should be performed in reverse order. run_mem_encoder (bool): Indicates if the memory encoder should be executed. prev_sam_mask_logits (torch.Tensor | None): Previous mask logits for the current object. inference_state (dict[str, Any], optional): The current inference state. If None, uses the instance's inference state. Returns: (dict): A dictionary containing the output of the tracking step, including updated features and predictions. Raises: AssertionError: If both `point_inputs` and `mask_inputs` are provided simultaneously. Notes: - The method assumes that `point_inputs` and `mask_inputs` are mutually exclusive. - The method retrieves image features using the `get_im_features` method. - The `maskmem_pos_enc` is assumed to be constant across frames, hence only one copy is stored. - The `fill_holes_in_mask_scores` function is commented out and currently unsupported due to CUDA extension requirements. r,r rlrrcurrent_vision_featscurrent_vision_pos_embedsrrsrtrirtrack_in_reverservrmaskmem_featuresrrrFrmaskmem_pos_enc) r_rprI track_steprEr=rrFr_get_maskmem_pos_enc)r1rirlrqrrrsrtrurvrr_rrrrrs r6r~z.SAM2VideoPredictor._run_single_frame_inference sP*AT-A-AFJFZFZ D !:G C7 #{':::jj++1!5&?!%##&|4$+!5,  ''9:  '.>.A.AmmDKKdkkFVFVZ`F`/B/K* +*.)B)B;O`Cacr)s %&r7c :|xs |j}|d}|yd|vr8t|tsJ|Dcgc]}|ddj}}||d<n|d}|djd}|dkDr!|Dcgc]}|j |ddd}}|Scc}wcc}w)aCache and manage the positional encoding for mask memory across frames and objects. This method optimizes storage by caching the positional encoding (`maskmem_pos_enc`) for mask memory, which is constant across frames and objects, thus reducing the amount of redundant information stored during an inference session. It checks if the positional encoding has already been cached; if not, it caches a slice of the provided encoding. If the batch size is greater than one, it expands the cached positional encoding to match the current batch size. Args: out_maskmem_pos_enc (list[torch.Tensor] | None): The positional encoding for mask memory. Should be a list of tensors or None. inference_state (dict[str, Any], optional): The current inference state. If None, uses the instance's inference state. Returns: (list[torch.Tensor]): The positional encoding for mask memory, either cached or expanded. Notes: - The method assumes that `out_maskmem_pos_enc` is a list of tensors or None. - Only a single object's slice is cached since the encoding is the same across objects. - The method checks if the positional encoding has already been cached in the session's constants. - If the batch size is greater than one, the cached encoding is expanded to fit the batch size. rNrrrr9)r_r<rclonerrr)r1out_maskmem_pos_encr_model_constantsr[rrqs r6rz'SAM2VideoPredictor._get_maskmem_pos_encns0*AT-A-A)+6  * 7!"5t<<<:M"NQ1Ra5;;="N"N5D 12"12C"D,Q/55a8JA~Q`&aAqxx BB'G&a#&a""#O'bs B4Bc |xs |j}t|d}|rdnd}ddtj|dg|jdd|j |j tj||jjfd|j |j tj|dfd |j |j d }t|D]}|d |} |d |} | |j|xs*| dj|xs| dj|} | |r|j||d ||dzr| d|d||dz| d |d ||dz|rytj|d|jdd} |jjr|jj!| } |j#|| d|d|\|d<|d<|S)aConsolidate per-object temporary outputs into a single output for all objects. This method combines the temporary outputs for each object on a given frame into a unified output. It fills in any missing objects either from the main output dictionary or leaves placeholders if they do not exist in the main output. Optionally, it can re-run the memory encoder after applying non-overlapping constraints to the object scores. Args: frame_idx (int): The index of the frame for which to consolidate outputs. is_cond (bool, optional): Indicates if the frame is considered a conditioning frame. run_mem_encoder (bool, optional): Specifies whether to run the memory encoder after consolidating the outputs. inference_state (dict[str, Any], optional): The current inference state. If None, uses the instance's inference state. Returns: (dict): A consolidated output dictionary containing the combined results for all objects. Notes: - The method initializes the consolidated output with placeholder values for missing objects. - It searches for outputs in both the temporary and main output dictionaries. - If `run_mem_encoder` is True, it applies non-overlapping constraints and re-runs the memory encoder. - The `maskmem_features` and `maskmem_pos_enc` are only populated when `run_mem_encoder` is True. rnrjrpNrrsize fill_valuerrF$@rrrobj_ptrobject_score_logitsrrrrrFrr$rTr)rqhigh_res_masksis_mask_from_ptsrr_rr)r_rWr=fullrCrrFrI hidden_dimrr_get_empty_mask_ptrrrrXnon_overlap_masks_for_mem_encr_run_memory_encoder) r1rlrrvr_rqrrrrroutrs r6rz6SAM2VideoPredictor._consolidate_temp_output_across_objsT>*AT-A-A9: .5*;S !%#** !=d&9&9!&<="&&{{ zz $**"7"78"&&{{  $):: !_ &&{{ $! 2Z( PG#23M#Nw#W -.CDWMO$[155i@L ##78<3 P8 ]] .ZZ# N zz77!%!N!N~!^X\XpXp%-!%$45J$K / YqY U / 02BCT2U r7c"|xs |j}|j|d\}}}|jj|d|||dt j ddg|j |j|ji|dddd }|d S) aGet a dummy object pointer based on an empty mask on the current frame. Args: frame_idx (int): The index of the current frame for which to generate the dummy object pointer. inference_state (dict[str, Any], optional): The current inference state. If None, uses the instance's inference state. Returns: (torch.Tensor): A tensor representing the dummy object pointer generated from the empty mask. r,TNrrrFrr) r_rprIrr=rrXrrF)r1rlr_rrrrs r6rz&SAM2VideoPredictor._get_empty_mask_ptrs*AT-A-AFJFZFZ[jko[pFqC7jj++#!5&?! Q$7DJJ$7t?O?OX\XcXcd&|4"!!%, 9%%r7cH|xs |j}|j|d|\}}}|jj|||||\} } |j | |} | j t j|j|jjdk(| fS)aRun the memory encoder on masks. This is usually after applying non-overlapping constraints to object scores. Since their scores changed, their memory also needs to be computed again with the memory encoder. Args: batch_size (int): The batch size for processing the frame. high_res_masks (torch.Tensor): High-resolution masks for which to compute the memory. object_score_logits (torch.Tensor): Logits representing the object scores. is_mask_from_pts (bool): Indicates if the mask is derived from point interactions. inference_state (dict[str, Any], optional): The current inference state. If None, uses the instance's inference state. Returns: maskmem_features (torch.Tensor): The encoded mask features. maskmem_pos_enc (torch.Tensor): The positional encoding. r,)rrpred_masks_high_resrrrr) r_rprI_encode_new_memoryrrEr=rrFr) r1rqrrrr_rrNrrrs r6rz&SAM2VideoPredictor._run_memory_encoder's2*AT-A-A.2.B.B?SWCXZd.e+a,0JJ,I,I!5! .- 3 -J- )/33O_U""-- $++BRBRV\B\#  r7ct|xs |j}|d}|t|tjsJ|d}|t|tsJ|dj D]Q\}}t ||dz} dd|d| |d| d} ||| | d<||D cgc]} | |  c} | d<| |||<Sycc} w) aISplit a multi-object output into per-object output slices and add them into Output_Dict_Per_Obj. The resulting slices share the same tensor storage. Args: frame_idx (int): The index of the current frame. current_out (dict): The current output dictionary containing multi-object outputs. storage_key (str): The key used to store the output in the per-object output dictionary. inference_state (dict[str, Any], optional): The current inference state. If None, uses the instance's inference state. rNrrrrr)rrrr)r_r<r=r>ritemsslice) r1rlrrr_rrrr obj_sliceobj_outr[s r6rz)SAM2VideoPredictor._add_output_per_objectQs*AT-A-A&'9:':6F +UUU%&78&*_d*KKK(78M(N(T(T(V > $G_gw{3I$(#'),7 B&y1)< G  +.>y.I*+*DS-Tqa l-T)*6=OK ( 3 >.Us B5cr|xs |j}|jj}|||jjzz }|||jjzz}t ||dzD]G}|ddj |d|dj D]}|dj |dIy)aRemove the non-conditioning memory around the input frame. When users provide correction clicks, the surrounding frames' non-conditioning memories can still contain outdated object appearance information and could confuse the model. This method clears those non-conditioning memories surrounding the interacted frame to avoid giving the model both old and new information about the object. Args: frame_idx (int): The index of the current frame where user interaction occurred. inference_state (dict[str, Any], optional): The current inference state. If None, uses the instance's inference state. rrirpNr)r_rImemory_temporal_stride_for_eval num_maskmemrrhr)r1rlr_rframe_idx_begin frame_idx_endtrs r6r}z3SAM2VideoPredictor._clear_non_cond_mem_around_inputts*AT-A-A JJ 6 6#a$***@*@&@@!A (>(>$>>  (9: GA M *+C D H HD Q#23H#I#P#P#R G 89==aF G Gr7c dj|d}||sdStd|dddtddk(rjdSt }|j d||j d ||D]}j ||d}ttt|jj|Dcgc]}|| } }ttt| } tt|  tt| | d<tt| | d <| d< fd } | d| d | d | d fd} | dd| dddScc}w)zRemove an object id from the tracking state. If strict is True, we check whether the object id actually exists and raise an error if it doesn't exist. rNrzCannot remove object id z/ as it doesn't exist. All existing object ids: rrrrrncg}D].}|j|}|vs|j||f0|j|yrb)rhrr&) containernew_kvskvold_idx_to_new_idx old_obj_indss r6 _map_keysz3SAM2VideoPredictor.remove_object.._map_keyssUG! ?MM!$**NN$6q$91#=> ?   W %r7rrc*||jD]w\}}|d|d<|dDcgc]}| c}|d<j|d|d<|d|d<|d|d<|d|d<j|||yycc}w)Nrrrrrr)rrr)rirrlrr[r_remain_old_obj_indsr1s r6 _slice_statez6SAM2VideoPredictor.remove_object.._slice_states"-k":"@"@"B j 3*-.@*ABU*V&'JMN_J`)aQ!,?*@)a%&)-)B)B3GXCY[j)k%&$' $56I$JL!!$Y0C!DI-01F-GH[-\)*++IsKYh+i j)as Brirjrp) rr|rWclear_all_points_in_videorr&clear_all_points_in_framerrcopyremover'r)r1r_rkstrictold_obj_idx_to_rmobj_input_frames_indsrl old_obj_idsold_idx new_obj_ids new_obj_indsrrrrrs`` @@@r6 remove_objectz SAM2VideoPredictor.remove_objects ,O<@@N  $&y11*6(3,,;I,F+GqJ  / 0A 5  * *? ;"9- -!$$$_5K%LM^%_`$$_5J%KL]%^_. OI  * *?Iv N O &i0 E#k"234 *//1""#45;NO{7+O OE#k"234 !#&9<"HI+/K0N+O(+/L+0N+O(%0 "  & /"89:/"789/"789/"<=> j _]35IJ_]35MNy))QPs9 Gc|j||}|d|j|d|d|j|d|d}||dj|d||dj|dt|d}d}t|D]}||d|vrd }n||d|vsd }n|s|d } |d } | dj || dj || dj|d} | | | d|<|d j|dt|D]*}|d |} | dj|d} | #| | d|<,t| ddk(r|j |yyy)zGRemove all input points or mask in a specific frame for a given object.rNrrrjrprnFTrirmrwrr)rrhrWrr_reset_tracking_results)r1r_rlrkrrrqframe_has_inputobj_idx2rirmrrrs r6rz,SAM2VideoPredictor.clear_all_points_in_frames%%fo> ./8<?CCItT )*BCGG SWX9: j) HO,BCHMM"&O,AB8LL"&  )-8K&56O&P # #$8 9 A A) L #$< = E Ei P2377 4HCDG 45i@ 89==iN!*- S"12G"H"R)*>?CCItT&KRO$<=iH  S;345:,,_=;)r7c0|j||dj|dj|dj|dj|dj|dj|djy) zCRemove all input points or mask in all frames throughout the video.rrnrrrrrN)r r)r1r_s r6rz,SAM2VideoPredictor.clear_all_points_in_video s $$_5(..0(..0 "((*./557-.446-.4462399;r7c|djD]}|j|djD]}|j|djD](}|dj|dj*|djD](}|dj|dj*|ddj|ddj|ddj|ddjd |d <|d jd |d <y )z8Reset all tracking inputs and results across the videos.rrrrjrprrirmFrrwNfirst_ann_frame_idx)rr)r_rs r6r z*SAM2VideoPredictor._reset_tracking_resultss`!!78??A A GGI  !67>>@ A GGI  !67>>@ 0A " # ) ) + & ' - - / 0!!;<CCE 0A " # ) ) + & ' - - / 0  &';<BBD &'?@FFH123GHNNP123KLRRT27./0177915-.r7c|jsy|xs |j}||jj|jjzz }|d}|dDcgc] }||ks | c}D]}|dj |d|j dijD]5}|dDcgc] }||ks | c}D]}|dj |d7ycc}wcc}w)z8Prune old non-conditioning frames to bound memory usage.Nrirpr)rer_rIrrrhrr)r1rlr_ min_framerirfrs r6rz)SAM2VideoPredictor._prune_non_cond_memory,s&& )AT-A-A 6 69c9c cc %m4 ))ABTa)m!T ?A 0 1 5 5a > ? /223H"MTTV GO!01I!J\AaR[ma\ G 89==aF G G U ]s C%C0 C$;C$r)r+)NNNrN)r_dict[str, Any] | NonerbrNN)FFNF)r1r2r3r4r r)rrnrrrzr{r5rdrrprr~rrrrrr}rrrr rr6r7s@r6r\r\<sF'$RV'&GmR415ji/jijiXH@H@TTT ''R4.4@"15L/L\'#X15 m / m ^&N26 ( / (V]a!>DY!>FG,L*L*\.>.>` < <66(Gr7r\ceZdZdZedddf d fd Ze d ddZddZe d ddZ ddZ dd Z dd Z d dd Z xZS)SAM2DynamicInteractivePredictoraSAM2DynamicInteractivePredictor extends SAM2Predictor to support dynamic interactions with video frames or a sequence of images. Attributes: memory_bank (list): OrderedDict: Stores the states of each image with prompts. obj_idx_set (set): A set to keep track of the object indices that have been added. obj_id_to_idx (OrderedDict): Maps object IDs to their corresponding indices. obj_idx_to_id (OrderedDict): Maps object indices to their corresponding IDs. Methods: get_model: Retrieves and configures the model with binarization enabled. inference: Performs inference on a single image with optional prompts and object IDs. postprocess: Post-processes the predictions to apply non-overlapping constraints if required. update_memory: Append the imgState to the memory_bank and update the memory for the model. track_step: Tracking step for the current image state to predict masks. get_maskmem_enc: Get memory and positional encoding from the memory bank. Examples: >>> predictor = SAM2DynamicInteractivePredictor(cfg=DEFAULT_CFG) >>> predictor(source=support_img1, bboxes=bboxes1, obj_ids=labels1, update_memory=True) >>> results1 = predictor(source=query_img1) >>> predictor(source=support_img2, bboxes=bboxes2, obj_ids=labels2, update_memory=True) >>> results2 = predictor(source=query_img2) Nr:ct||||d|_g|_t |_t tt|x|_ |_ ||_ y)aInitialize the predictor with configuration and optional overrides. This constructor initializes the SAM2DynamicInteractivePredictor with a given configuration, applies any specified overrides Args: cfg (Any): Configuration dictionary containing default settings. overrides (dict[str, Any] | None): Dictionary of values to override default configuration. max_obj_num (int): Maximum number of objects to track. Default is 3. This is set to keep fixed feature size for the model. _callbacks (dict | None): Dictionary of callback functions to customize behavior. TN) r(r)r` memory_bankr obj_idx_setrrrrrn _max_obj_num)r1r2r3 max_obj_numr4r5s r6r)z(SAM2DynamicInteractivePredictor.__init__Zs[& i4!%52=ikHZ>[2\\T/'r7c|j||j|j|jddjdd||||\}}}|rt |t r|g}|Jd| |Jd|stjt|ddf|j|j}tjt|dftj|j}|t|t|k(sJd t|t|k(sJd |j|||||j}|d |d } } t|jdk(r t!d t#|j} | | | | } } tj$| dz d} | j'dd| j'ddfS)aoPerform inference on a single image with optional bounding boxes, masks, points and object IDs. It has two modes: one is to run inference on a single image without updating the memory, and the other is to update the memory with the provided prompts and object IDs. When update_memory is True, it will update the memory with the provided prompts and obj_ids. When update_memory is False, it will only run inference on the provided image without updating the memory. Args: im (torch.Tensor | np.ndarray): The input image tensor or numpy array. bboxes (list[list[float]] | None): Optional list of bounding boxes to update the memory. masks (torch.Tensor | np.ndarray | None): Optional masks to update the memory. points (list[list[float]] | None): Optional list of points to update the memory, each point is [x, y]. labels (list[int] | None): Optional list of labels for point prompts (>0 for positive, 0 for negative). obj_ids (list[int] | None): Optional list of object IDs corresponding to the prompts. update_memory (bool): Flag to indicate whether to update the memory with new objects. Returns: res_masks (torch.Tensor): The output masks in shape (C, H, W) object_score_logits (torch.Tensor): Quality scores for each mask rrNr;)rrr^r]r`r_z3obj_ids must be provided when update_memory is TruezDbboxes, masks, or points must be provided when update_memory is Truerz,masks and obj_ids must have the same length.z-points and obj_ids must have the same length.rrzMNo objects have been added to the state. Please add objects before inference.r)r)rprqrXr%rrr<rBr=rrWrrFr update_memoryrrr|rrr~) r1r,r]r_r^r`rrrrrrs r6rnz)SAM2DynamicInteractivePredictor.inferenceys< R $ 5 5jjjjmA&,,Ra0 !6!  '3'")& ](] ]&$(: V :~c'lAq%9AQAQZ^ZeZefc'lA%6ekkRVR]R]^ 5zS\1a3aa1v;#g,. _0_ _.   w >oo' "-l";[I^=_K t A %no o4##$",S/;s3CK ll;#3; !!!Q')<) 1 QR@RS"&&{{  zz'')>)>?"&&{{  $)::''+&&{{ $ 2#7+ pIAvD--- --))#f+6G     )!1#;s 5E','85!:d#dD$(8 c:c c8//'5%>C|,~~bc*.>|.L.R.RSUSV.WW*+;L+I+O+OPRPS+T*UU^_g_m_mnpnq_r^ssAH@IIJKWIQ .w1EEH^ +GgkB(C/UXYnUo$%:;GgPQkR% p( \ * - -dkk HXHX\bHb - c   :: 3 3!ZZJJ>ZN,0JJ,I,I!%!2!2 . 01F G! -J- )/0@+,.=*+  01r7ct|jdk(st|tr'|jd|j j z}nM|j\}}|j j|jdd|jdd||d}|jdddj|j|j jjg|jdS)aPrepare memory-conditioned features for the current image state. If ``obj_idx`` is provided, features are prepared for a specific prompted object in the image. If ``obj_idx`` is None, features are prepared for all objects. If no memory is available, a no-memory embedding is added to the current vision features. Otherwise, memory from previous frames is used to condition the current vision features via a transformer attention mechanism. Args: obj_idx (int | None): The index of the object for which to prepare the features. Returns: pix_feat_with_mem (torch.Tensor): The memory-conditioned pixel features. rr9N)currcurr_posmemory memory_posnum_obj_ptr_tokensrr;)rWrr<rBrOrIrKget_maskmem_encmemory_attentionr rLrrd_modelr)r1rpix_feat_with_memr*memory_pos_embeds r6$_prepare_memory_conditioned_featureszDSAM2DynamicInteractivePredictor._prepare_memory_conditioned_featuress t A %GS)A!% 1 1" 5 8O8O O (,';';'= $F$ $ ; ;&&rs+//4+#$ !<! 7 ((Aq166    JJ ' ' / / __R   r7cgg}}|jD]}|j|djdjddd|ddjdjddd}||jj |jj dz z}|j|tj|d}tj|d}||fS)zfGet memory and positional encoding from memory, which is used to condition the current image features.rr;rrrr9r) rrr~rLrImaskmem_tpos_encrr=r)r1 to_cat_memoryto_cat_memory_pos_embedr maskmem_encr*r1s r6r-z/SAM2DynamicInteractivePredictor.get_maskmem_enc@s13R. $ 0 0 8   !12D!E!M!Ma!P!X!XYZ\]_`!a b*+<=bAII!LTTUVXY[\]K% (C(CDJJDZDZ]^D^(__K # * *; 7  8 =a0 99%,026+/#'$(#@C %@C)@C0 @C ) @C ! @C"@C@C +@C@CD% %)&*&*%) N2!N2$N2$ N2 # N2  N2N2`! F ( 4#%)%)$( 1 1 #1 # 1 " 1  1 r7rc6eZdZdZgdZdZdfd ZdZxZS) SAM3PredictorzSSegment Anything Model 3 (SAM3) Interactive Predictor for image segmentation tasks.) rN)rO)HrPc>t|||tjgdj dddj |j |_tjgdj dddj |j |_y)zTSetup the SAM3 model with appropriate mean and standard deviation for preprocessing.)_@rSrSr9rN) r(rr=rrrErFrGrH)r1rIrr5s r6rzSAM3Predictor.setup_modelsp E7+LL!67<HHK"^3<HHK"^3<dZd Zd!Z d?d"Z d@d#Z dAd$Z dB dCd%Z! dDd&Z"e dE dFd'Z# dG dHd(Z$dId)Z% dJd*Z& dKd+Z' dLd,Z( dMd-Z)dNd.Z*d/Z+ dOd0Z,d;dPd1Z-d2Z.ed3Z/xZ0S)QSAM3VideoSemanticPredictorz9Segment Anything Model 3 (SAM3) Video Semantic Predictor.皙?$r9irr;rMrQNrArr: TFrc&t||||||_||_||_||_||_| dkDr| | ksJ| | ksJ| |_| |_| |_ | |_ | |_ ||_ ||_ ||_||_||_||_d|_d}d}||_||_||_||_||_||_||_t5||_i|_|j:dj=|j>y)zTInitialize the SAM3VideoSemanticPredictor with configuration and optional overrides.rNi'r)r3r^) r(r)score_threshold_detectiondet_nms_threshassoc_iou_threshtrk_assoc_iou_threshnew_det_threshhotstart_delayhotstart_unmatch_threshhotstart_dup_threshinit_trk_keep_alivemax_trk_keep_alivemin_trk_keep_alive8suppress_overlapping_based_on_recent_occlusion_thresholdsuppress_det_close_to_boundary*decrease_trk_keep_alive_for_empty_maskletso2o_matching_masklets_enablefill_hole_area _dist_pg_cpumax_num_objectsnum_obj_for_compilerecondition_every_nth_framemasklet_confirmation_enable+masklet_confirmation_consecutive_det_threshreconstruction_bbox_iou_threshreconstruction_bbox_det_scorertrackerr_rcrrd)r1r2r3r4rrrrrrrrrrrrrrrrrrrrrrrr5s r6r)z#SAM3VideoSemanticPredictor.__init__ s@\ i4)B&, 0$8!, A *n< <<&.8 88,'>$#6 #6 "4"4 D E/M+:d7,H),  .#6 +F(+F(;f8.L+-J**I> ! )*11$//Br7ct|||ddlm}||jj d}|j j|dy)z+Setup the SAM3VideoSemanticPredictor model.rrVF) with_backbone)rIrN)r(rrZrWr*rIr)r1rIrrWr5s r6rz&SAM3VideoSemanticPredictor.setup_model sB E7+6'tyyeL   ue >OO]](X&ss C,3"C' C,'C,ct|jdkDry|jJ|jjdk(sJ|jj}|giddg|zd}||_y)aInitialize an inference state for the predictor. This function sets up the initial state required for performing inference on video data. It includes initializing various dictionaries and ordered dictionaries that will store inputs, outputs, and other metadata relevant to the tracking process. Args: predictor (SAM3VideoSemanticPredictor): The predictor object for which to initialize the state. rNr)rtracker_inference_statestracker_metadata text_promptper_frame_geometric_prompt)rWr_r r$r)rrr_s r6rdz%SAM3VideoSemanticPredictor.init_state" s~ y(( )A -   ,,,  %%000&&-- $(* "+/&:*=  %4 !r7c|jjdz }||jd<d|jvr|j|||||j |dS)z.^ s/h1A/hrrrobj_id_to_tracker_scorer)background_valuer)#sortedrr<rr rrIrrWr=rrFrrrrLrrrrr*ranyrr'rrrBmax._apply_object_wise_non_overlapping_constraintsr rrr%rr )r1rrrr curr_obj_idsrrrwrkpred_idsrrrtracker_scoresrr_rvrrs r6rz&SAM3VideoSemanticPredictor.postprocessC s]/0n1134 )T*55i@I$(JJ$4$4$@   b |  !%)5;;vdkk+R J#VvN6$:#V\]^Jz'7'7'9$'?1ASASTVUVAW^hijklorrJ||L JL]L]^H,,@LMf()&1MV`VgVgK||>JKF'/KT^TeTeH $))..0JNNvN4NND#D)J,Z8JXd^AtG4k$6G 6RT\]aTbclTmntvJ"ii/hjQRTUQUFVFZFZF\F`F`FbefFf@g/h&h!i"Q&!&'3  # &/H)II"";Y[k"l,0LL,>,>,a,a #- )[[ - 1  KK (8 9 r7c |j|||| } |ik(r|j|j|j|||\} } |j ||| | | ||\} }| j dt }|j||| || }|j| | || |}|d}|d}tj|d | d d }| jd d kDrM| jj} |d }|d|jtt|| ||||||| fS)aThis function handles one-step inference for the DenseTracking model in an SPMD manner. At a high-level, all GPUs execute the same function calls as if it's done on a single GPU, while under the hood, some function calls involve distributed computation based on sharded SAM2 states. - `input_batch` contains image and other inputs on the entire video; it should be identical across GPUs - `tracker_states_local` holds the local masklet information in this GPU shard - `tracker_metadata_prev` manages the metadata for SAM2 objects, such as which masklet is hold on which GPUs it contains both global and local masklet information )r,rlrmr)rlrr)rlrudet_outtracker_low_res_masks_globaltracker_obj_scores_globalrrreconditioned_obj_ids)rlrrrtracker_update_plan)rrrrrrrnum_objnum_obj_dropped_due_to_limit)num_obj_trackednum_obj_droppedrrr)run_backbone_and_detectionr&_initialize_metadatarun_tracker_propagation!run_tracker_update_planning_phaserr"run_tracker_update_execution_phase build_outputsr?rrrr}rr'r)r1r,rlrlrrurmrrrrrrrrrrrrrrtracker_obj_idss r6rz/SAM3VideoSemanticPredictor._det_track_one_frame s.11-!5 2  !B & ! ( ()B)B)D EBFB^B^!5"7C_C ?$&?594Z4Z)E&?"7!55[5 11!4 7 78OQTQV W$(#J#J!!5 3 $K$  ++)E"7 3"7 , //@A,_=  "vv&:9&EF23QR % * *1 - 1(A(I(I(K(R(R(T %3I>O !E Fy Q X XS*CDE     $  %  r7c|jd\}}}}||zdz }||zdz }||kD|d|z kz||kDz|d|z kz}|S)zSuppress detections too close to image edges (for normalized boxes). boxes: (N, 4) in xyxy format, normalized [0,1] margin: fraction of image r9r;r)unbind) rvmarginx_miny_minx_maxy_maxx_cy_crs r6&_suppress_detections_close_to_boundaryzASAM3VideoSemanticPredictor._suppress_detections_close_to_boundaryQ sk&+\\"%5"ueUu}!u}!f sV|!34f EsU[|I[\ r7c|j|}|jj|||}|j||}|j ||S)z.Run backbone and detection for a single frame.rk)rprIrq_extract_detection_outputs_cache_backbone_features)r1r,rlrmrr-sam3_image_outrs r6rz5SAM3VideoSemanticPredictor.run_backbone_and_detection_ s]''+55!HGW6 11.BVW %%n5r7cn|djdj}|s|dz }tjt t |j d|j|jdddfj|}|d}|d}||jkD}||||||||d S) z%Extract and filter detection outputs.rxr9gחArrNpred_boxes_xyxyr)bboxr$rr) rr}r=rrrrrrrFr~r)r1r r pred_probsrrrrs r6r z5SAM3VideoSemanticPredictor._extract_detection_outputsk s#M2::2>FFH ##c)J<< z''*+ ,""$$  T' 9Z(  )):;#L1 D:::#D)t$ &D>   r7c|jjj}|dd}|j|dd|j |dd|ddg}|d|d|d }||j_y ) z'Build and cache SAM2 backbone features.rMsam2_backbone_out backbone_fpnrrr;r9vision_pos_enc)vision_featuresrrN)rrIrWconv_s0conv_s1rM)r1r rWrRtracker_backbone_fpntracker_backbone_outs r6r z3SAM3VideoSemanticPredictor._cache_backbone_features s<<-->>~./BC  $ $U>%:1%= >  $ $U>%:1%= > . !! $  4B7#$450  %9 !r7c|j||\}}}tj||dk(sJdj||d|}|}||fS)zGRun the tracker propagation phase for a single frame in an SPMD manner.rlrz{} != {})&_propogate_tracker_one_frame_local_gpur?rir) r1rlrr obj_ids_locallow_res_masks_localobj_scores_locallow_res_masks_globalobj_scores_globals r6rz2SAM3VideoSemanticPredictor.run_tracker_propagation s@D?j?j I@k@ < *,<vvm'}(|j5d })|(j5d }*tA|)|*d }+|+|j&ks| |jBk\s.d}| jE|C|jFd kDxr"||jFzd k(xrt|d kD},|,s|r|jI|||||||jKd }-|-d kDrQt!|dr |j"r%|jLdkDr|jO||||||}|jQ||||d}.t|d kDrtjR|.|g}.t|d kDr%tjT|.t+|}/|.|/}.|.|d<t|.|d<t|d kDr|djWtY|| ||djWtY|| ||d|jWtY|| |t[|dtjZ||d<|D]*}0d|d|0<d|d||0<|d j]|0d,d|vsJ|j^r$|ja|d|d|d||!}1|1|d<||fS#t.$rYvwxYw)"zKRun the tracker update planning phase for a single frame in an SPMD manner.rrrrr max_obj_id)rrrrrobj_id_to_last_occludedr3r$rrr) det_masks det_scores_np trk_masks trk_obj_idsrzhitting self.max_num_objects=z with new_det_num=z and prev_obj_num=rr_warm_up_complete)rlrudet_to_matched_trk_obj_idsnew_det_obj_idsempty_trk_obj_idsunmatched_trk_obj_idsr)new_det_fa_indsr;r=r:obj_ids_newly_removedrr(rFrNr;r:rTr)r?gr4)robj_ids_all_gpu_prevobj_ids_all_gpu_updatedr:r;)1rrrLrr_associate_det_trkrr r?rrWrr warning_drop_new_det_with_obj_limitrhasattrr9_process_hotstartrrrindex ValueErrorr&rr rrrr=rrFrrr#rr1rr/_suppress_overlapping_based_on_recent_occlusion_tracker_update_memories concatenateisinr&rrrhr"update_masklet_confirmation_status)2r1rlrurrrrrrrdet_mask_predsr6 det_cls_np det_bbox_xyxyr>r=r:r(r<r prev_obj_num new_det_numrnew_det_num_to_keepr; metadata_newr?rshould_recondition_iour)r*det_box det_scoretrk_idx tracker_mask mask_binary mask_areatracker_box_pixels mask_height mask_widthtracker_box_normalized det_box_batchtracker_box_batchrshould_recondition_periodicrqupdated_obj_ids_this_gpu is_removedrkrs2 r6rzO(PQ%&;O&LM2:;PQu;v2w')"#8#FG !$(/v$+H$5$;$;$=$A$A$C$I$I$K !(!5!5!7!;!;!=!C!C!E &-fo   # #$'2-i8 $   ! & +   . .>>}_?]^D-dhhj.>.>.@AOvv3I>? /* '($ + %(<(< < NN;d&:&:%<> >>o.K0 =ABIIkDZZ  5j AB t01T5K5K262H2H#+E /"3&;%3I3 / !<%(E !+7Z( /.%:*D%:,H/N%:  "'  . . 2s;Z7[^_7_'F'L'L'N$ :# G!&/'2#H-g6 "#8#CDJJ:VG &99N9Nq9Q%R%Z%Z[\%]"*6*<*?B<<<dNpNpAp-1*)--j9I$ :N  , ,q 0 9D<<<A 934q8 $ '*@  & &/$%)  266q9 >4!459O9OPPSVV373g3g!4-,- 40  ) )*> Yu ) v$8 #B  ! #')~~7OQ`6a'b $ $ % )!94@U;VWJ'? 'L $*BY'*-.F*GY'  ! # !2 3 : :3P]^mPn;o p  1 8 8_jYhNi9j k !E Fy Q X XO]?%CD 255I,5WY[Y_Y_`oYp1q  ., NF>B !2 3F ;\` !E Fy QRX Y !: ; ? ? M N 1111  + +>>-j9%:9%E(~&*4 1#n2E1Fd:,W 4"'#1 **CDHH  7=EZ7Z 3 3`d`t`t#F#M#M"'**   " "WW3" K!D G GH G UVO,;k,I ) 2 8 8 : ?H ; <YbbpXq?ETWf)'GaK@@? !: ; 9=8I8I ( 5++K <?s A%EEc|d}|d}|}|} |d} t| dkDr3tj| } |d| } |j|||| |}t| dkDr|j || |S)zEExecute the tracker update plan for a single frame in an SPMD manner.r>r;r?rr$)rlrr new_obj_masksr)rWr=rD_tracker_add_new_objects_tracker_remove_objects) r1rlrrrrr>r;new_det_obj_ids_localnew_det_fa_inds_localr?new_det_fa_inds_local_t new_det_maskss r6rz=SAM3VideoSemanticPredictor.run_tracker_update_execution_phase s'::K&L&9:K&L,;,;*=>U*V $ % )&+&6&67L&M #*1&/:Q*RM#'#@#@#%1+%9 $A$  $ % )  ( ()=?T U##r7c|d}|d}i}|d}|jd} t|t| k(sJt|| D] \} } | || < tj|} |d| jd} t|t| k(sJt|| D] \} } | || < |Ut|dkDrG|j di}|D]0} |j | }||d|jd|| <2|S)z-Build the output masks for the current frame.r>r;rrr$rr()r rWrr=rDr)rrrrrr>r;rexisting_masklet_obj_idsexisting_masklet_binaryrkr$new_det_fa_inds_tnew_det_low_res_masksr(r*s r6rz(SAM3VideoSemanticPredictor.build_outputs sa'::K&L&9:K&L$9#C ">"H"H"K+,4K0LLLL 8:QR *LFD%)N6 " *",,_= '0A B L LQ O?#s+@'AAAA1FG *LFD%)N6 " * ! ,5J1Ka1O.A.E.EFgik.l +/ S9==fE&-4V_W-E-O-OPQ-RN6*  S r7c$|jtjk(sJd|jtj|j d|j tj}t |dkr|St|jd|jd}||jk\}tj|d} |jd} |jd} |stjntj} | | | | z| dkDz} | | | | z| dkDz}| jd|jdz}tj dr|| j#j%} |j#j%}|j#j%}| j&d}t)|D]J}t)|D]:}| ||fs tj*d |d ||d ||d ||d || <Lt)|D]J}t)|D]:}|||fs tj*d |d ||d ||d ||d || <L|S) NzExpected boolean tensor, got rrr)diagonalr9rrz frame_idx=z: Suppressing obj z last occluded z in favor of )rr=rErrrFrWrr~rtriur gtltrr isEnabledForrrrrrr')r1binary_low_res_masks last_occludedrrlrurormask_iou_threshoverlapping_pairslast_occ_expanded_ilast_occ_expanded_jcmp_opsuppress_i_masksuppress_j_maskrqrfjs r6rjzSSAM3VideoSemanticPredictor._get_objects_to_suppress_based_on_most_recently_occluded8 s"$))UZZ7u;XYmYsYsXt9uu7kk % %a ('..** w<1  +33A68L8T8TUV8WX!^!^^!JJC+55a8+55a8!(ehh (*=> ?"R' )  (*=> ?"R' )  &))a)0?3F3F13F3MM    r "y'<-11399;O-11399;O)--/557M)..q1J:& z*A&q!t, )yl*O?9-.!3=A\\=\=\9>]> :K*Nk40 00   -  % %&7&?&?&B C  " ">#9#9!#< = > ! "Q &"')),>A"F $yya@ "5"="=a"@ "'++a"]$2E2Ea2H"]QUQ\Q\"] ${{1T[[A 13CCCr7c  |j}|j}|j}|jsJd|jsJd|j dt |k(s%Jd|j ddt ||j ddk(rwt j|j d}t jgt j} t jgt j} i} i} || | | | fS|j ddk(rkt jgt j}|dkDjdjj} || } || } i} i} || | | | fS|jdd|jddk7rt j|jddt j|jddkrEtj |j#d |jddd d j%d }nDtj |j#d |jddd d j%d }|dkD}|dkD}t'|j)d j+|j)d j+}|jj}|j,redd lm}d |z }||\}}t j2|j dt4}t7||D]\}}|||f|k\sd||<n||k\jd}|jdjj} t j8| |}||} || } t j8||k\t j:t j||k\d }t j<|d}i} i} t j>|d }||j@k\|z}t jB|d |jDk\}tGt j<||zd}tI|j dD]4}|||ddf|k\| |<||vs|||jK}|| |<6|| | | | fS)ajMatch detections on the current frame with the existing masklets. Args: det_masks: (N, H, W) tensor of predicted masks det_scores_np: (N,) array of detection scores trk_masks: (M, H, W) tensor of track masks trk_obj_ids: (M,) array of object IDs corresponding to trk_masks Returns: new_det_fa_inds: array of new object indices. unmatched_trk_obj_ids: array of existing masklet object IDs that are not matched to any detections on this frame (for unmatched, we only count masklets with >0 area) det_to_matched_trk_obj_ids: dict[int, np.ndarray]: mapping from detector's detection indices to the list of matched tracklet object IDs empty_trk_obj_ids: array of existing masklet object IDs with zero area in SAM2 prediction z'float tensor expected (do not binarize)rz7trk_masks and trk_obj_ids should have the same length, rfrxrrNrrFr)linear_sum_assignmentrTr)&rrris_floating_pointrrWr?rrint64rrrrrprodrrr rrr~rLrscipy.optimizerrrEr logical_and logical_notnonzeroargmaxr-rHIGH_IOU_THRESHrrr&) r1r5r6r7r8 iou_thresholdiou_threshold_trkrr>r=r<r:r(trk_is_nonemptydet_masks_binarytrk_masks_binaryiousious_npr cost_matrixrow_indcol_indtrk_is_matcheddrtrk_is_unmatched is_new_detdet_to_max_iou_trk_idxdet_is_high_confdet_is_high_ioudet_is_high_conf_and_iour)s r6rBz-SAM3VideoSemanticPredictor._associate_det_trk s.--  55,,**,W.WW,**,W.WW,~~a C $44 EinnUVFWEXX\]`al]m\n o 4 >>!  ! ii q(9:O$&HHR$: ! "RXX 6 )+ &.0 +%*/!  ^^A ! # hhr2884O(1}11f1=AACIIKO$/$@ ! +_,< = )+ &.0 +%*/!   ??23 9??23#7 7wwyrs+,rwwyrs7K/LLMM''*"-#"'  '!* MM''*"-#"'  '!* %q=$q=(00399;=M=U=UVW=X=^=^=`a((*""$  , , <g+K4[A GWXXinnQ&7tDNGW- -11a4=$55(,N1% -&)::??Q?GN*..6.:>>@FFH>>/N?K +,< ='(89^^ ^ + NN266']":C D **Z03&("*,'!#7!;)T-B-BBzkQ&&q1T5I5II#&rzz2B_2T'UVW'X#Y y~~a() @A,71 8V,W &q ),,()?)BCHHJ >?/ ;  @  ! & +    r7c |d|d}|d} |d} |d} t} |s||jz n||jz} |D] }|vr||<|| vsJ|j| |<"t}|jD]}|j ||D]!}t |j | |dz| |<#|D]5}||j|t|j| |dz | |<7|jr&|D]!}t|j| |dz | |<#|jD]\}}|| vs|| vrt||jk\rL|| kDxr| xs || kxr|}|r/| j|tj d|d|d || |d ks||| vs|| vstj d|d|d | j||jD][\}}t|d kr|st |fd nt|fd}|D] }||k7s ||f}| |j|"]| jD]r\\}}}|| vs|| vr|| kDr|r || ks%|s(t||j"k\sA| j|tj d|d|d|d|t| j | | |fS)zEHandle hotstart heuristics to remove unmatched or duplicated objects.obj_first_frame_idxunmatched_frame_indstrk_keep_aliveoverlap_pair_to_frame_indsrrzRemoving object r$z# since it is unmatched for frames: rz, due to being unmatchedr;c|Srbrcr[rs r6z>SAM3VideoSemanticPredictor._process_hotstart..qs 7J17Mr7)keyc|Srbrcrs r6rz>SAM3VideoSemanticPredictor._process_hotstart..ss?!"23%-.J%K""#45 #?F D$7$77IX\XkXkLk & >F00.7#F+/ //%)%=%=N6 "  > u $>$E$E$G 6    4 5 6" ^F%()@)@.QWBX[\B\%]N6 " ^, ^F  ( / / :&))@)@.QWBX[\B\%]N6 "  ^  : :+ b),T-D-DnU[F\_`F`)av& b&:%?%?%A 2 !FM(F6K,K=!T%A%AA&9&&AM&Q&aZaVa&'/-?KG#&)--f5LL*6(*YKH==JOM v&!+/1"77 /xz)Ldef%))&1) 20'A&F&F&H F "A"&'!+ '-MN,2RS . F00.7C.s3::9E F F"6P5U5U5W  1 "\6M(F6K,K#F+m;G#F+m;}%)A)AA)--f5LL*6(*YKH@@L~\ZgYhj  45$h..r7ct|dk(rytj|jd|jdd}t |dr |j r%|jjj|}tj|dkDjd d d }d}|}|D]}t|d } | dk(r|| z} ||| } ||| } | jd} |jj| | | d| }|\}}|d}dD]S}|||vr ||||d<|Dcgc]}|c}|||d<|jj||||||U|| z }ycc}w)zHRun Sam2 memory encoder, enforcing non-overlapping constraints globally.rNrrFrr9rgrrrr)rr_rirrr)r_rlrr)rWrrr rrEr9rrI"_suppress_object_pw_area_shrinkager=rrrrr)r1rrlr?rr start_idx_gpustart_idx_state tracker_statenum_obj_per_state end_idx_statelocal_high_res_maskslocal_object_score_logitslocal_batch_size encoded_memlocal_maskmem_featureslocal_maskmem_pos_encrirposs r6rJz3SAM3VideoSemanticPredictor._tracker_update_memoriess ' (A -   # #A &##  t01T5K5K!\\//RRSabN#kk>A+=*B*Bx*B*PRVX]^ '5" 1M #M)$< =  A%+.??M#1/-#P (;OM(Z %388; ,,:: $)!& - ;K=H 9 "$9' 6KO  K $<<J` K(34FGYnIoRU#Io K(34EF 33$1' +K 8 C + 4  0 0OE" 12Jps, E0cVt|dkDr|dnd}|jj|}d|d<||jddnd|d<t||j dk(sJ|j sJt j|jd|jddjd}|dkD}t||D]'\}} |jj|||| d  )|jj||j||S) z*Add a new object to SAM2 inference states.rN)rr,rMrFrrr%)rWrrrrrrrr rrrrzr{r) r1rlrrrtrprev_tracker_statenew_tracker_state new_obj_idr+s r6ruz3SAM3VideoSemanticPredictor._tracker_add_new_objectssO9<4 8`d .);=#5#5a#8888..000  # #A &##  '!*  &) %( ]$C  J LL ( ( 1#!z* )   112CD##$56##r7c|syg}|D]J}|D] }|jj||d"t|ddkDs:|j|L||ddy)zgRemove an object from SAM2 inference states. This would remove the object from all frames in the video.NFrrr)rrrWr)r1rr active_statesstaterks r6rvz2SAM3VideoSemanticPredictor._tracker_remove_objectssw  ) ,E! H **5&*G H 5#$q($$U+ ,#0Qr7ctjgtjtjdtjdiit t id}it t t tt t td}|jrLtjgtjtjgtjd|d<||d<|S)z%Initialize metadata for the masklets.rr9)rrr3rrrr4)rrrrr)rconsecutive_det_numrr) r?rrrrr'rrBrrr)r1rrs r6rz/SAM3VideoSemanticPredictor._initialize_metadatasxxBHH-xx288,!2=d2C') $&$/$5)#.*5d*;"u    + +((2rxx0(*xxBHH'= 0H+ ,(0$r7c(|d}|d}|d}|j|jk(s!Jd|jd|jt|D cic]\} } | |  } } } tj||} || } tj| D cgc]} | |  c} tj }|j }tj||}|| ||<tj|}|| ||<tj||}|jD]}|tj||z}tj||dzd }||jk\}|j||<||d<||d<|Scc} } wcc} w) zZUpdate the confirmation status of masklets based on the current frame's detection results.rrrzGot rfr)rrr) rrrr?rLrrr full_like zeros_likerrr CONFIRMED)r1rr@rAr:r;confirmation_data status_prevconsecutive_det_num_prevrrkobj_id_to_updated_idxprev_elem_is_in_updatedprev_elem_obj_ids_in_updatedprev_elem_inds_in_updatedunconfirmed_valrr is_matchedrchange_to_confirmeds r6rMz=SAM3VideoSemanticPredictor.update_masklet_confirmation_status(s%%;<(1 #45J#K   $8$>$>> ;$$%T*>*D*D)E F >AJJa@b cf c c"$''*>@W"X';r6 num_to_keep score_orders r6rDz7SAM3VideoSemanticPredictor._drop_new_det_with_obj_limitcs K73#777777 ! 88B) ) #o. ." "jj!?@2F )+l{*CDr7r)r/rr)FNr+)r)T)r,rIrlrIrlrBrrBrurErmrr list[Any]rrJrrE)g?)r,rIrlrIrmrrrE)rlrBrrrdict[str, np.ndarray]) rdict[str, torch.Tensor]r( list[int]rrrrrrI)rlrBrurErrrrIrrIrrrrr) rlrBrrIrrJrrJr?zset[int]rurE) rlrBrrBrrrrrrrb) rrrrIrrrrrz set | None)NF) rrIrrrrrlrHrurE)rrrlrB)r5rIr6 np.ndarrayr7rIr8r)rlrBrurEr:dict[int, np.ndarray]r;rr<rr=rrrJ)rrrlrBr?rI) rlrBrrBrrrtrIrr)rrrr) rrJr@rrArr:rr;r)rrrrE)1r1r2r3r4r-rrkrhrirrrCrr r)rr r5rdrnrr~rrrrr rr r rr1rrIrrrjrrBrFrJrurvrrMrrrDr6r7s@r6rr sCOLNOKIN F "&#&! !AD37%*',$&$(56'*&)WVC VCp=^440F3j4l %%N4&*i i i  i  i  i !i (i  .i #i V    *6 JP hl  .9"774=7Vk7,($)($*3 ($ ( ($ 0 ($$0($TR9R9R9) R9 '3 R9 $0 R9 5R9(R9vD,D,'3D, . D, - D, ( D,D,L"$"$"$) "$ ( "$ 3 "$H -1 $($&2$ 5$3 $ * $$V!% B*B!B B  B  BHD8C C "C  C  C Jm/m/m/%: m/ $ m/ & m/ *m/!m/^81(181>A81R^81t*$*$*$ *$ $ *$ ( *$X0$" H. .).", . %: . $ .`W3  r7r)6r4 __future__r collectionsrrrrtypingrrrr?r=torch.nn.functionalnn functionalrultralytics.data.augmentrultralytics.engine.predictorr ultralytics.engine.resultsr ultralytics.utilsr r r ultralytics.utils.metricsrrultralytics.utils.torch_utilsrramgrrrrrrrrrsam3.geometry_encodersrrr9r\rrLr]rrrcr7r6rs#0  .6.667M   +@ ' @ 'FKCIKC\AGAGH I mI X RMR0J MJ Z3/+]3/lJ!6Jr7