CO3Dv2

[3D挑战中的常见对象]((https:// eval.ai/web/challenges-page/1819/overview)，允许在隐藏的测试服务器上进行透明评估-挑战自述文件中的更多详细信息
序列数量增加2倍，帧数增加4倍
改善图像质量-由于更好的视频解码而减少了块状伪影
改进的分割蒙版-稳定跟踪主前景对象而不跳转到背景对象
允许下载约100序列的较小的单序列子集，该序列仅由用于评估多视图单序列任务的序列组成
数据集文件托管在20 GB的块中，有助于更稳定的下载
一种新颖、更用户友好的数据集格式
序列中的所有图像都被裁剪为相同的高度x宽度

Common Objects in 3D Challenge (https://eval.ai/web/challenges-page/1819/overview) enables transparent evaluation on a hidden test server; more details can be found in the challenge README. The number of sequences has been doubled, and the total frame count has been quadrupled. Image quality has been improved, with reduced blocking artifacts achieved through better video decoding. Segmentation masks have been enhanced to support stable tracking of the primary foreground object without drifting to background elements. A smaller single-sequence subset comprising approximately 100 sequences is available for download, with these sequences exclusively used for evaluating multi-view single-sequence tasks. Dataset files are hosted in 20 GB chunks to facilitate more stable downloads. A novel, more user-friendly dataset format has been adopted. All images within each sequence are cropped to identical height × width dimensions.

搜集汇总

背景与挑战

背景概述

CO3Dv2是一个用于3D对象重建挑战的公开数据集，由Facebook AI Research和伦敦大学学院于2021年发布。该数据集在CO3D基础上进行了显著增强，包括序列数量增加2倍、帧数增加4倍，并改善了图像质量和分割蒙版，以支持更稳定的多视图评估。它提供了一种新颖的用户友好格式，图像被统一裁剪，并托管在20GB的块中以便下载。