WildDet3D-Stereo4D-Bench
收藏资源简介:
WildDet3D Stereo4D Benchmark 是一个用于3D物体检测的基准数据集,包含来自Stereo4D视频的真实立体深度信息。数据集从Stereo4D测试集中抽取了每段视频的1帧(约占所有Stereo4D视频的7%)。数据规模为1K到10K之间,具体包含383张验证集图像和2,782个标注。测试集不公开,用于隐藏评估。数据集目录结构包括annotations(包含Stereo4D_val.json、Stereo4D_test.json和Stereo4D_val_class_map.json)、images(7,704张512x512图像)、depth(7,704个.npy深度图,float32格式,单位为米)和camera_parameters(7,704个.json相机内参)。标注格式采用COCO3D,包含图像元数据、3D边界框(中心点、尺寸、旋转矩阵、3D/2D边界框等)和类别列表。数据集适用于3D物体检测任务,遵循CC BY-NC许可,仅供研究和教育使用。
WildDet3D Stereo4D Benchmark is a benchmark dataset for 3D object detection, containing real stereo depth information from Stereo4D videos. It extracts 1 frame from each video in the Stereo4D test set, accounting for approximately 7% of all Stereo4D videos. The dataset scale ranges from 1K to 10K, specifically including 383 validation set images and 2,782 annotations. The test set is not publicly available and is used for blind evaluation. The dataset directory structure includes annotations (containing "Stereo4D_val.json", "Stereo4D_test.json" and "Stereo4D_val_class_map.json"), images (7,704 512×512 images), depth (7,704 .npy depth maps in float32 format with unit of meters) and camera_parameters (7,704 .json files of camera intrinsic parameters). The annotation format adopts COCO3D, which includes image metadata, 3D bounding boxes (center point, dimensions, rotation matrix, 3D/2D bounding boxes, etc.) and category lists. This dataset is applicable to 3D object detection tasks and follows the CC BY-NC license for research and educational use only.




