Training with 360 deg annotations and 360 deg inference

[bmankirLinker]

Since KITTI dataset 3d bbox labels are the ones only on cam_2, in some parts of code there are checks for valid gt by checking if it resides in the image (due to GT-AUG) such as https://github.com/sshaoshuai/PointRCNN/blob/master/lib/datasets/kitti_rcnn_dataset.py#L267 However, after commenting out that valid point selection, providing 360 labeled data, and extending the scope with negative z-coordinate in cam I still can't get 360 inference results. Only getting the front objects. Recently going through the Pointnet part, would be glad if anyone has any experience with such modification. Thanks

[bmankirLinker]

For anyone interested, the distance_based_proposal method https://github.com/sshaoshuai/PointRCNN/blob/master/lib/rpn/proposal_layer.py#L58 has to be updated. Such that negative values should be added to nms_range_list since it checks z in the cam dimension. I was able to get behind (negative z)detections afterward.

[DaHaiHuha]

Hi, thanks for your comment, i'd wish i have seen your post earier, Great job! Could you make a PR to share your contributions? Although i think make another mask as codes suggested should work @bmankirLinker

[DaHaiHuha]

Update: just as @bmankirLinker said, add negative values to nms_range_list and it worked, thanks!

[pyaf]

Hey, what value are you using for nms_range_list? nms_range_list = [-40.0, 40.0, 80.0] should work, right?

[bmankirLinker]

Hey, what value are you using for nms_range_list? nms_range_list = [-40.0, 40.0, 80.0] should work, right?

Depends on the z range you want to define, it indeed works, but you won't get the proposals further than -40 on the negative side.

[hulsmeier]

@bmankirLinker Hey, can i ask what 360 dataset you used?