构造 image A region + image B region 的跨图方向 QA（复用 object cross 逻辑）。

构造 image A object + image B region 的跨图方向 QA（复用 object cross 逻辑）。

构造"给定 obj1 相对 obj2 的方向，问 Region 3 相对 obj2 的方向"QA（MCQ）。 复用 object proxy 主逻辑，并在输出语义上将 C 明确为 region。

构造"站在物体A，面向物体B，问region C的方位和距离"QA（MCQ格式）。 复用 object relpos 主逻辑，并在输出语义上将 C 明确为 region。

Build SAM3 image model Args: bpe_path: Path to the BPE tokenizer vocabulary device: Device to load the model on ('cuda' or '

Build SAM3 dense tracking model. Args: checkpoint_path: Optional path to checkpoint file bpe_path: Path to the BPE tokenizer

Loads an image, sends it with a text prompt to the service, saves the results, and renders the visualization.

调用vLLM生成3D instance的描述 注意：描述应该是3D一致的，不包含2D图像中的位置信息

Returns canonicalized `text` (lowercase and punctuation removed). From: https://github.com/google-research/big_vision/blob/53f18caf27a9419231bbf08

Concatenates a list of RotatedBoxes into a single RotatedBoxes Arguments: boxes_list (list[RotatedBoxes]) Retur

Concatenates a list of BitMasks into a single BitMasks Arguments: bitmasks_list (list[BitMasks]) Returns:

axis_method: - "zclose": choose floor axis (among its 3 axes) closest to world_up - "largest_face": choose normal of largest face ==

Clean up debug files when function successfully returns

Clear detector added conditioning frame if it is within a predefined window of the refined frame. This allow model to update masks on these fr

Close a session. This method is idempotent and can be called multiple times on the same "session_id".

Add new points and labels to previous point inputs (add at the end).

Concatenate rel pos coeffs to the q & k tensors, so that qk^T is now effectively including rel pos biases. Args: q (Tensor): q te

Computes connected components labeling on a batch of 2D integer tensors using Triton. Args: input_tensor (torch.Tensor): A BxHxW int

Function that recursively copies data to a torch.device. Args: data: The data to copy to device device: The device to which the d

Count the total number of images present in the messages history.

解码RLE格式的mask

Returns a default color palette. Returns: ColorPalette: A ColorPalette instance with default colors. Example:

从深度图生成点云 Args: depth: 深度图 (N, H, W) intrinsics: 内参矩阵 (N, 3, 3) extrinsics: 外参矩阵 (N, 4, 4) - world to camera

将整幅深度图投影到世界坐标，同时可返回颜色。 Returns: pts: (N,3) float32 cols: (N,3) uint8 或 None

将整幅深度图投影到世界坐标，同时可返回颜色。 Returns: pts: (N,3) float32 cols: (N,3) uint8 或 None

Draws keypoints of an instance and follows the rules for keypoint connections to draw lines between appropriate keypoints. This follo

Args: circle_coord (list(int) or tuple(int)): contains the x and y coordinates of the center of the circle.

Draw panoptic prediction annotations or results. Args: panoptic_seg (Tensor): of shape (height, width) where the values

Draw a rotated box with label on its top-left corner. Args: rotated_box (tuple): a tuple containing (cnt_x, cnt_y, w, h,

Draw semantic segmentation predictions/labels. Args: sem_seg (Tensor or ndarray): the segmentation of shape (H, W).

将 mask 编码为 RLE 格式，与 agent.py 中的保存方式一致

将 mask 编码为 RLE 格式，与 agent.py 中的保存方式一致

尽量兼容大小写和扩展名差异寻找图片路径。

Ensure output is a directory (create if needed).

Ensure output is a directory (create if needed).

Evaluate spherical harmonics at unit directions using hardcoded SH polynomials. Works with torch/np/jnp. ... Can be 0 or more batch d

从 QA entry 中提取需要判别唯一性的引用（Ref）。

从mask路径中提取场景ID 例如: sam_masks_debug/0a7cc12c0e/DSC05865/bed.png -> 0a7cc12c0e

用中位数+MAD 计算逐轴 Z 分数，删除 |Z| 最大的 10% 点。 这样不依赖固定阈值，按比例裁掉尾部离群。

Return the color that is the farthest from the given color. Args: img_array (np array): any *x3 np array, 3 is the RGB c

优先从 mask_index.json 获取 mask_rle；如果缺失（比如某些场景 mask_index.json 没有 mask_rle 字段）， 则回退到 bbox json 中该 instance 的 mask_encodings（与 highest_confidence

根据mask路径和场景ID找到原图路径

Forward pass of the backbone combiner. :param samples: The input images :param captions: The input captions :param input_boxe

Process multi-level features through the transformer encoder. Args: src: List of multi-level features, each with shape (

This method is only used for benchmark eval (not used in the demo).

Args: input: NCHW images rois: Bx5 boxes. First column is the index into N. The other 4 columns are xyxy.

Get the image feature on the input batch.

Compute the detector's detection outputs in a distributed manner, where all GPUs process a chunk of frames (equal to the number of GP

为QA条目生成所需的图片，返回图片路径列表和元数据。 Returns: (image_paths, metadata): 图片路径列表（相对于output_images_dir）和元数据

Generate masks by comparing rendered depth with mesh depth

A generic version of `torchvision.ops.nms` that takes a pairwise IoU matrix.

A generic version of `torchvision.ops.nms` that takes a pairwise IoU matrix. (CPU implementation based on https://github.com/jwyang/faster-rc

Get 1D sine positional embedding as in the original Transformer paper.