PointRCNN with Centerness loss (FCOS:Fully Convolutional One-Stage Object Detection on 3D) for robust 3D object detection. (mAP 95.12% on KITTI-Easy vs 85.95%)[To be released]
Along with Argoverse dataloader.
Trained seperate for Pedestrain and Vehicle classes.
Vanilla AB3DMOT modified for PointRCNN output.
Used mahalanobis distance and feature.
Note: The repo incudes Tracker with MLP refinement in run_ab3dmot_mod.py (Needs to be complete)
Pipline is Detection -> Tracker -> MLP-refine -> IDs, Locations, Size
Without Groundtuth (Colored Trackers) & With Groundtuth (white + Colored Trackers):
- PointRCNN with Centerness loss + Non-NMS regression
- PointRCNN with Autoregressive Transformer regression
- PointRCNN with PointCNN w. knn-graph Backbone (Performs better in RPN, more backbones can be tried)
- PointRCNN with MeteorNet Tracker
- Tracker with MLP
- Tracker with LSTM
- Tracker with PointNet local features (Points inside BBOX)
- Fusion with stereo-images(360) and then using Frustum pointnet with 2D+3D ground-truth (mAP 96.48% on KITTI-Easy)
- Fusion with range-image and PointGNN