MegaDepth-X (MD-X)
收藏资源简介:
MegaDepth-X是由康奈尔大学和哈佛大学联合构建的大规模3D重建数据集,作为MegaDepth的升级版,其规模扩大7倍至1865个场景、44万张图像。该数据集通过互联网照片集合构建,经过严格的动态内容过滤、多视角立体匹配优化以及单目深度引导的深度图修复,最终生成高精度稠密深度信息。其创新性在于采用MASt3R-SfM框架解决视觉相似场景误匹配问题,并引入稀疏感知采样策略模拟长尾分布场景特性,旨在提升3D基础模型在稀疏、噪声图像条件下的重建鲁棒性,特别针对对称场景和重复结构等挑战性场景。
MegaDepth-X is a large-scale 3D reconstruction dataset jointly developed by Cornell University and Harvard University. As an upgraded version of the original MegaDepth dataset, its scale has been expanded 7-fold to 1865 scenes and 440,000 images. Constructed from Internet photo collections, this dataset undergoes rigorous dynamic content filtering, multi-view stereo matching optimization, and monocular depth-guided depth map inpainting, ultimately generating high-precision dense depth information. Its core innovation lies in adopting the MASt3R-SfM framework to resolve mismatching problems in visually similar scenes, and introducing a sparse-aware sampling strategy to simulate the characteristics of long-tail distribution scenarios, aiming to enhance the reconstruction robustness of 3D foundation models under sparse and noisy image conditions, particularly for challenging scenarios such as symmetric scenes and repetitive structures.

- 1Long-tail Internet photo reconstruction康奈尔大学; 哈佛大学·肯普纳研究所 · 2026年



