Propagate the prompts to get grounding results for the entire video. This method is a generator and yields inference outputs for all

Args: rgb (bool): whether to return RGB colors or BGR colors. maximum (int): either 255 or 1 Returns: ndarray: a vec

Revert `inference_state` to what it was right after initialization.

Unified visualization function. If zoom_in_index is None: - Render all masks in input_json (equivalent to visualize_masks_from_resul

Args: img_size (int): Input image size. Only relevant for rel pos or rope. patch_size (int): Patch size.

Args: classes (list[int] or None): scores (list[float] or None): class_names (list[str] or None): is_crowd (list[

Create SAM3 visual backbone with ViT and neck.

Compute the image features on a given frame.

`maskmem_pos_enc` is the same across frames and objects, so we cache it as a constant in the inference session to reduce session stor

从 images 路径列表中推断 mask_index.json 的路径。 Args: scene_id: 场景 ID images: mask 图像路径列表（来自 bbox_json 的 images 字段） Returns:

Randomly modifies given color to produce a slightly different color than the color given. Args: color (tuple[double]): a

Load and resize an image and convert it into a PyTorch tensor.

从 images 路径列表中推断并加载 mask_index.json 文件。 Args: scene_id: 场景 ID images: mask 图像路径列表（来自 bbox_json 的 images 字段） Returns:

判断世界坐标点是否在 region bbox 内（优先 OBB，失败则退化到 AABB）。

Run tracking on a single frame based on current inputs and previous memory.

Add text, point or box prompts on a single frame. This method returns the inference outputs only on the prompted frame. Note

构造"image A 的物体1 + image B 的物体2，问物体2在image B的方位"QA（MCQ）。

构造"给定 obj1 相对 obj2 的方向，问 obj3 相对 obj2 的方向"QA（MCQ）。 使用代理坐标系（由 obj1->obj2 与 assumed_dir 确定），仅输出 N/E/S/W。

计算 OBB（定向包围盒），返回 (transform, extents)。 Returns: transform: 4x4 变换矩阵（从 OBB 局部坐标系到世界坐标系） extents: (3,) OBB 的尺寸

Args: binary_mask (ndarray): numpy array of shape (H, W), where H is the image height and W is the image width. E

Args: segment: numpy array of shape Nx2, containing all the points in the polygon. color: color of the polygon. Refer

从mask key中提取A/B/C/1/2标签.

Return an activation function given a string

将 PIL Image 转换为 base64 编码的字符串。

将 PIL Image 转换为 base64 编码的字符串。

The inverse function for sigmoid activation function. Note: It might face numberical issues with fp16 small eps.

compute bounding box given an input mask Inputs: - masks: [B, 1, H, W] tensor Returns: - box_coords: [B, 1, 4], contains (x, y)

Fallback: estimate center, axes, extents from the 8 corners using PCA. PCA on the 8 corners of a box recovers its axes (up to sign/permutati

Returns (center(3,), R(3,3), extents_full(3,)) if available, else None.

Render a two-panel visualization with a cropped original view (left/upper) and a zoomed-in mask overlay (right/lower), then return it as a PI

Args: img: same as in __init__

Encodes a collection of masks in RLE format This function emulates the behavior of the COCO API's encode function, but is executed partially

从图像对的两个图像中分别采样所有不同的对象对。返回 [(image_pair, (inst1, inst2)), ...] inst1来自图像1, inst2来自图像2 Args: max_inst_pairs_per_image_pair: 每个 imag

Sends a request to the OpenAI-compatible API endpoint using the OpenAI client library. Args: server_url (str): The base URL of the s

Map angle to [-180, 180).

Distribute the new objects to the GPUs with the least workload.

计算3D bbox的轴对齐包围盒（AABB）。 Returns: (min_corner, max_corner) 两个3D点，表示AABB的最小和最大角点

Compute detection outputs on a chunk of frames and store their results in multigpu_buffer.

构建 stem -> instances 映射，过滤 ignore_labels。

Remove the non-conditioning memory around the input frame. When users provide correction clicks, the surrounding frames' non-conditio

Create geometry encoder with all its components.

Create a grayscale version of the original image. The colors in masked area, if given, will be kept.

Create position encoding for visual backbone.

Create SAM3 transformer encoder and decoder.

Create segmentation head with pixel decoder.

Create SAM3 text encoder.