search

prstrive / UniMVSNet

[CVPR 2022] Rethinking Depth Estimation for Multi-View Stereo: A Unified Representation
MIT License
228 stars 12 forks source link

Different results between released model and released ply #15

Open xy-guo opened 2 years ago

xy-guo commented 2 years ago

I tried to infer and fuse on DTU dataset, but found the predicted results are different from your released ply models.

For example: Your released point cloud for scan1 (mvsnet001_l3.ply): 26727801 points, fscore 0.191357
Predicted point cloud for scan1: 19405933 points, fscore 0.2162721

When I visualize the two point clouds, the completeness of predicted point cloud is much lower than released point cloud, especially around edges.

My environment: Fusibile: compiled with cuda 11.4, sm86 Pytorch: 1.8.2+cu111 Python: 3.8.12

Here is my log file for scan1 log.txt

xy-guo commented 2 years ago

I tried to set align_corners=True for grid_sample, and I got more points (22287197), but still less than 26727801.

prstrive commented 2 years ago

Maybe you can try Pytorch1.2.0, cudatoolkit 10.0 and python3.6.

cainsmile commented 2 years ago

I tried to infer and fuse on DTU dataset, but found the predicted results are different from your released ply models.

For example: Your released point cloud for scan1 (mvsnet001_l3.ply): 26727801 points, fscore 0.191357 Predicted point cloud for scan1: 19405933 points, fscore 0.2162721

When I visualize the two point clouds, the completeness of predicted point cloud is much lower than released point cloud, especially around edges.

My environment: Fusibile: compiled with cuda 11.4, sm86 Pytorch: 1.8.2+cu111 Python: 3.8.12

Here is my log file for scan1 log.txt

Hi, I used gipuma to fuse the depth maps, but I the number I got is only half as the author provided. So if there is any parameters of setting I need to change?