计算 OBB 的 8 个角点（世界坐标）。 Args: transform: 4x4 变换矩阵 extents: (3,) OBB 尺寸 Returns: corners: (8, 3) 8 个角点的世界坐标

计算 OBB 的 8 个角点（世界坐标）。 Args: transform: 4x4 变换矩阵 extents: (3,) OBB 尺寸 Returns: corners: (8, 3) 8 个角点的世界坐标

Given two lists of boxes of size N and M, compute the intersection area between __all__ N x M pairs of boxes. The box order must be (xmin

Args: polygons (list[ndarray]): each array has shape (Nx2,) height, width (int) Returns: ndarray: a bool mask of sha

处理单个JSON文件，为每个instance生成描述

Propagate the added prompts to get grounding results on all video frames.

Remove an object from tracking.

Resets the image embeddings and other state variables.

从图像对的两个图像的实例并集中采样3个不同的对象。返回 [(image_pair, (instA, instB, instC)), ...] Args: max_triplets_per_pair: 每个图像对最多生成的3元组数量，超过则随机采样

在 sample_pairs 基础上增加 region 约束的图像对采样。 当 at_least_3_objs_1_region=True 时，要求： - 图像对中的 object 实例并集数量 >= 3； - 图像对中的 region 实例并集数量 >= 1。

Sets the image on which we want to do predictions.

Sets the text prompt and run the inference

Expects a torch tensor with length 2 in the last dimension. The coordinates can be in absolute image or normalized coordinates, If th

hotstart_delay: int, the delay (in #frames) before the model starts to yield output, 0 to disable hotstart delay. hotstart_unmatch_th

Encodes prompts for input to SAM's mask decoder. Arguments: embed_dim (int): The prompts' embedding dimension im

Applies non-overlapping constraints object wise (i.e. only one object can claim the overlapping region)

Match detections on the current frame with the existing masklets. Args: - det_masks: (N, H, W) tensor of predicted masks

构造单选选项，随机放置正确答案。

构建以A为原点、面向B的局部坐标系。 返回: (forward, right, up) 三个单位向量，构成右手坐标系。 forward: 从A指向B的方向 right: 右手坐标系中的右方向 up: 上方向

Build SAM-style prompt encoder and mask decoder.

Depending on the brightness_factor, gives a lighter or darker color i.e. a color with less or more saturation than the original color. A

Compile all model components for faster inference.

Compile the SAM model with torch.compile for speedup.

Construct an initial `BatchedDatapoint` instance as input.

Convert different format of boxes to an NxB array, where B = 4 or 5 is the box dimension.

Convert a low-res mask to video resolution, matching the format expected by _build_tracker_output. Args: low_res_mask: T

Convert different format of masks or polygons to a tuple of masks and polygons. Returns: list[GenericMask]:

Corner order used by the provided JSONs: 0 (-,-,-) 1 (-,-,+) 2 (-,+,-) 3 (-,+,+) 4 (+,-,-) 5 (+,-,+) 6

Create dot product scoring module.

Create the SAM3 image model.

Create the SAM3 Tracker memory encoder.

Create the SAM3 Tracker transformer components.

Create transformer decoder with its layer.

Create transformer encoder with its layer.

Create ViT backbone for visual feature extraction.

Create ViT neck for feature pyramid.

Create visual-language backbone.

This function handles one-step inference for the DenseTracking model in an SPMD manner. At a high-level, all GPUs execute the same fu

Find proper places to draw text given a box.

Drop a few new detections based on the maximum number of objects. We drop new objects based on their detection scores, keeping the hi

When outputting a single mask, if the stability score from the current single-mask output (based on output token 0) falls below a thr

Embeds box prompts.

Embeds mask inputs.

Embeds point prompts.

Find a free port (a random free port from 1024 to 65535 will be selected) https://stackoverflow.com/questions/1365265/on-localhost-ho

Gets the batch size of the output given the batch size of the input prompts.

Get a dummy object pointer based on an empty mask on the current frame.

Compute stability scores of the mask logits based on the IoU between upper and lower thresholds.