mattia-durso/Uni10K
收藏资源简介:
Uni10K是一个大型户外重建数据集,覆盖了约100,000平方米的大学校园场景。该数据集包含从地面和空中视角拍摄的10,000多张4K分辨率图像。由于图像匹配和映射的复杂性,标准的COLMAP重建在此规模下通常需要数周的计算时间。为了使重建可行,使用了空间和时间先验以及从粗到精的重建方案。首先,利用GPS信息将图像匹配的复杂性从二次降低到近线性,通过限制图像比较在预定义的空间半径内。其次,由于帧是从视频序列中采样的,最初从每秒1帧的均匀采样帧中重建基线模型。然后通过增量注册、三角测量和使用局部束调整细化附加图像来增加相机覆盖密度。该过程以几轮全局束调整结束。最终重建包含10,000多张图像、620万个稀疏点,整体平均重投影误差约为0.63像素。训练在全分辨率图像上进行,每第8帧按字母顺序保留作为测试集的一部分。
Uni10K is a large-scale outdoor reconstruction dataset of a university campus covering approximately 100,000 m². The scene is captured from both ground-level and aerial viewpoints and contains more than 10,000 4K-resolution images. A standard COLMAP reconstruction at this scale would typically require weeks of computation because of the complexity of image matching and mapping. To make reconstruction feasible, we use spatial and temporal priors together with a coarse-to-fine reconstruction scheme. First, GPS information is used to reduce the image-matching complexity from quadratic to near-linear by limiting image comparisons to a predefined spatial radius. Second, because the frames are sampled from video sequences, we initially reconstruct a baseline model from frames sampled uniformly at 1 frame per second. Camera coverage is then densified by incrementally registering, triangulating, and refining additional images using local bundle adjustment. The process concludes with several rounds of global bundle adjustment. The final reconstruction contains more than 10,000 images, 6.2 million sparse points, and an overall mean reprojection error of approximately 0.63 pixels. Training is performed on the full-resolution images. Every 8th frame, sorted alphabetically, is held out as part of the test set.




