search

huang-yh / SelfOcc

[CVPR 2024] SelfOcc: Self-Supervised Vision-Based 3D Occupancy Prediction
Apache License 2.0
273 stars 17 forks source link

The reprojection loss does not decrease during training. #17

Open Fanqyu opened 2 months ago

Fanqyu commented 2 months ago

Hi,

Thanks for your inspiring work.

Recently, when I used your code on the MVS reprojection loss, I found that it doesn't go down during training. However, according to the test results, it does affect the performance. Specifically, when the total amount of data is small, it seems to improve performance, but when the data amount increases, the performance does not improve as expected.

Is this normal?

Looking forward to your reply. It'll help me a lot.

Best regards.

Liu202209 commented 2 weeks ago

Hello, I would like to ask if you have reproduced the results of nusecenes_occ? The results of my direct training are very poor. If you have successfully reproduced it, would you please let me to see your environment files to see if it is my environment problem.