The AcfNet test results on KITTI2015
ziming-liu opened this issue · comments
Helllo, I have tried to reproduce the results of Acfnet in the paper.
1- I use the scene-flow epoch20-long pretrained parameters for initialization.
2- I train the acfnet-uniform for 600epoch on 2 GTX 2080Ti with 2X2batch size, train the acfnet-adaptive for 275 epoches on 2 GXT2080Ti with 2X1 batch size (lilimited by the GPUmemory)
But the test results from official website are much poor than the results in the paper. I directly use the given config for kitti2015 in this repo.
What problem maybe cause this??? anyone can give an idea?
acfnet-uniform result:
Error | D1-bg | D1-fg | D1-all |
---|---|---|---|
All / All | 1.96 | 4.92 | 2.45 |
All / Est | 1.96 | 4.92 | 2.45 |
Noc / All | 1.78 | 4.59 | 2.24 |
Noc / Est | 1.78 | 4.59 | 2.24 |
acfnet-adaptive result:
Error | D1-bg | D1-fg | D1-all |
---|---|---|---|
All / All | 2.04 | 4.08 | 2.38 |
All / Est | 2.04 | 4.08 | 2.38 |
Noc / All | 1.87 | 3.79 | 2.18 |
Noc / Est | 1.87 | 3.79 | 2.18 |
the result in the paper:
ALL-D1-ALL 1.89
NOC-D1-ALL 1.72
Hi, @ziming-liu
1st, make sure you pull the latest code as we fixed a bug about 2 months ago.
2nd, make sure you use the same config and training schedule as we do because KITTI is kind of hard to get a good result. As you only have 2 GPUS, you can try our published config first and then try the same iterations as our training schedule.
3rd, you can try our released 'LocalSoftArgmin' code, as we found it gives a more stable result and sometimes maybe better.
Please tell me the result you have tried, and give more details, such as tensorboard and conf, disp map for analysing your problems.