- Currently support training only proposal stage.
- Still in beta stage but you are welcome to try -- may have bugs.
- No multi-GPU support at the moment -- but training on KITTI is fast.
- I assume velodyne points have already been filtered to camera field of view. I will release code soon to carry out this filtering. (As a crude approximation you can filter the points with a cone from +45 to +135 degrees.)
- Support only KITTI dataset at the moment. I have tried to keep code general to allow adding support for new datasets.
Some comments:
- Note that these results are cherry-picked.
- Trained for 11 epochs using
car.yamlmodel. - Training proposal layer takes about 10 minutes per epoch.
Learning rate seems to need tuning (observe classification loss).
