Test on stereo dataset

[darthandvader]

你好，请问如果我的数据集是stereo在同一camera pose的情况下采集的请问有什么建议吗，如果用组合成hypernerf的格式貌似并不可行。

[darthandvader]

主要是在运行bash colmap.sh的时候貌似由于单一camera pose，colmap会运行不成功，导致没有points3D_downsample.ply文件，无法训练，请问是否有绕开这一步的方法呢？似乎dnerf是不需要额外执行colmap的

[pengcanon]

Were your videos captured with a stereo camera? Suppose image frames from each camera are stored in a different folder, then you may run colmap separately on both folders with camera pose fixed for each folder (but obviously different between the two). Dnerf has camera poses labelled and stored up in json files already but it still needs something like colmap to generate the point3D.ply file.

That being said, I'm also unable to train anything other than dnerf because I can't seem to get the images, point clouds results, camera poses in a trainable format so that I can run train.py on them. Hopefully the author will provide a more detailed guidance on this.

[pengcanon]

Sorry I take back what I just said. You shouldn't run colmap per folder. Instead, run colmap per pair of image frames from the two folders. Don't fix anything as the two cameras are likely different. With two images, I think colmap is still able to generate point clouds of good quality.

[darthandvader]

as for point clouds I think I can extract them for each frame separately, but for hypernerf format dataset there seems to be just one fused.ply file after being processed by colmap. Any instruction on how which ply file shoud I use for downsample? Thanks!

[guanjunwu]

Hi, you can also customize the size of bounding box, and generate the random point in the bounding box. Endo-GS seems successfully solved it. Because the dataset used only has 2 static cameras.