Calculate absolute positional embeddings. If needed, resize embeddings and remove cls_token dimension for the original embeddings. Ar

Dispatch a request based on its type.

Dispatch a request based on its type.

Dispatch a stream request based on its type.

Dispatch a stream request based on its type.

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

根据 scannetppv2 路径推断对应的 DptV3 depth_da3 npy 路径。 例: /.../scannetppv2/data/<scene>/dslr/resized_undistorted_images/IMG.JPG -> /.../DptV3

根据 dense 路径推断对应的 depth_da3 npy 路径。 例: /.../dense/rgb/IMG.png -> /.../dense/depth_da3/IMG.npy

根据 scannet 路径推断对应的 depth_da3 npy 路径。 例: /.../scannet/scans_test/scene0707_00/color/00000.jpg -> /.../scannet/scans_test/scene0707_00/

根据 scannetppv2 路径推断对应的 DptV3 depth_da3 npy 路径。 例: /.../scannetppv2/data/<scene>/dslr/resized_undistorted_images/IMG.JPG -> /.../DptV3

Equivalent to nn.functional.interpolate, but with support for empty channel sizes.

从 COLMAP 数据中加载相机 pose 和内参 Args: colmap_dir: COLMAP sparse 重建目录路径 (例如: "proxy-gs/berlin/sparse/0") Returns: imag

Load the covisibility matrix; return None if missing or ambiguous.

Load the covisibility matrix; return None if missing or ambiguous.

加载 da3 深度图，替代高斯渲染。 DA3 预处理会在 scene_dir / "depth_da3" 下保存每帧深度；兼容旧的 dense/depth_da3。

Load a dummy video with random frames for testing and compilation warmup purposes.

Load an image as a single-frame video.

Load video frames from either a video or an image (as a single-frame video). Alternatively, if input is a list of PIL images, convert its for

从 ScanNet v2 格式场景中加载相机 pose 和内参。 目录结构示例: scannetv2/scans/scene0000_00/ color/ # RGB 图像 cam/ # 每帧一

读取 nerfstudio transforms_undistorted.json，返回 intr_map[stem] = 3x3 K c2w_map[stem] = 4x4 camera-to-world scene_type: str s

读取相机与图像路径，返回 intr_map[stem] = 3x3 K c2w_map[stem] = 4x4 camera-to-world scene_type: str（scannetppv2 / dl3dv / scannet） si

Load intrinsics/extrinsics and frame list for a DL3DV scene.

Load intrinsics/extrinsics and frame list for a DL3DV scene.

Load intrinsics/extrinsics and frame list for a scannetppv2 scene.

Load intrinsics/extrinsics and frame list for a scannetppv2 scene.

Load intrinsics/extrinsics and frame list for a ScanNetV2 scene. Scene structure: data_root/scene_id/ with color, intrinsic, pose folders. E.g

Load the video frames from video_path. The frames are resized to image_size as in the model and are loaded to GPU if offload_video_to_cpu=Fal

Load the video frames from video_path. The frames are resized to image_size as in the model and are loaded to GPU if offload_video_to_cpu=Fal

Load the video frames from a directory of image files ("<frame_index>.<img_ext>" format)

Load the video frames from a directory of JPEG files ("<frame_index>.jpg" format). The frames are resized to image_size x image_size and are

Load the video frames from a video file.

Load the video frames from a video file.

Load video from path, convert to normalized tensor with specified preprocessing Args: video_path: Path to video file image_s

初始化 vLLM 服务客户端。 Args: api_url: vLLM 服务的 API 地址，默认为 DEFAULT_VLLM_API_URL Returns: api_url: 返回 API URL 字符串（保持接口兼容性）

初始化 vLLM 服务客户端。 Args: api_url: vLLM 服务的 API 地址，默认为 DEFAULT_VLLM_API_URL Returns: api_url: 返回 API URL 字符串（保持接口兼容性）

将点云与 3D bbox 发往 rerun 0.24.1 进行可视化。

将点云与 3D bbox 发往 rerun 0.24.1 进行可视化。

Randomly pick 3 distractors from remaining 7, but if yaw is close to the sector boundary, forbid the adjacent sector on that boundary sid

将 mask 内像素 + 深度投影到世界坐标，返回 (N,3)。

将 mask 内像素 + 深度投影到世界坐标，返回 (N,3)。