Get a string for PyTorch and GPU properties (for logging and debugging).

Handle the case of visual prompt. Currently, in the inference API we do not explicitly distinguish between initial box as visual prom

Identical to _init_box. Except that C=2 for points (vs. 4 for boxes).

Encode keypoint locations into a target heatmap for use in SoftmaxWithLoss across space. Maps keypoints from the half-open interval [x1, x2)

Load model checkpoint from file.

Compute the image backbone features on the fly for the given img_ids.

Fuse the current frame's visual feature map with previous memory.

Handle hotstart heuristics to remove unmatched or duplicated objects.

Return a new messages list that contains only: 1) messages[:2] (with optional warning text added to the second message's content) 2) the lates

Setup model device and evaluation mode.

Args: Same as in :meth:`__init__()`. Returns: fig (matplotlib.pyplot.figure): top level container for all th

Enable TensorFloat-32 for Ampere GPUs if available.

Start worker processes for handling model inference.

Suppress detections too close to image edges (for normalized boxes). boxes: (N, 4) in xyxy format, normalized [0,1] margin:

This function suppresses masks that shrink in area after applying pixelwise non-overlapping constriants. Note that the final output c

Suppress overlapping masks based on the most recent occlusion information. If an object is removed by hotstart, we always suppress it if it o

Remove an object from SAM2 inference states. This would remove the object from all frames in the video.

Run Sam2 memory encoder, enforcing non-overlapping constraints globally.

Directly turn binary `mask_inputs` into a output mask logits without using SAM. (same input and output shapes as in _forward_sam_head

Whether to use multimask output in the SAM head.

Deprecated method. Please use `add_new_points_or_box` instead.

Add new points to a frame.

Add text, box and/or point prompt on a specific video frame.

Add a new point prompt to Tracker. Suppporting instance refinement to existing objects by passing existing obj_id or adding a new object by pa

Given a text prompt and an image, this tool will perform all aspects of agentic problem solving, while saving sam3 and MLLM outputs to their

Build a new OBB orientation R_new whose z-axis is aligned to up_axis, with minimal rotation. extent_mode: - "keep": keep original s

将mask应用到图像上，用红色轮廓圈出物体（只绘制轮廓，不填充）

Split N items into chunks <= max_chunk, but as even as possible. Returns a list of (start, end) index ranges.

将 OBB (transform, extents) 转成 8 个角点的 JSON 格式。

将 OBB (transform, extents) 转成 8 个角点的 JSON 格式。

Batched version of box_iou. Boxes should be in [x0, y0, x1, y1] format. Inputs: - boxes1: Tensor of shape (..., N, 4) - boxes2: Tens

Camera rotation MCQ: - Keep your verified R_rel and YXZ as_euler. - Choose two most significant axes among yaw/pitch/roll. - De

Exact distance question (numeric answer, meters).

Thresholded distance yes/no question with auto/semi-random threshold near the true distance.

Main-direction template for camera translation QA.

Returns: - stem -> label -> list[(ins_id, description)] - stem -> ins_id -> {"label": ..., "description": ...}

基于 bbox json 与 mask 路径构造“给定 A 视角的标记，问物体在 B 视角相对方向”QA。

构造"给定 A 视角中的物体 1 和 B 视角中的物体 2，问它们之间的 3D 距离是多少"的 QA。

基于 bbox json 与 mask 路径构造“给定 A 视角的 mask，问物体在 B 视角距离多少”QA。 json_path: 3D bbox 结果文件，单元素格式示例： { "ins_id": "1", "label": "b

构造"给定 A 视角的物体1 和 B 视角的物体2，问哪个更高/更高(看顶部)"的 QA（MCQ）。