July 2020
tl;dr: CenterPoint with bells and whistles wins the 2020 Waymo Open Dataset challenge.
Overall impression
The anchor-free object detection of AFDet is very close to CenterPoint.
Key ideas
- Anchor free, NMS free. 1000x faster on CPU. Embedded system friendly.
- Architecture
- PointPillars backbone.
- The backbone only has two stage network and keeps the same high resolution for object detection. No FPN needed as in BEV only one scale is needed.
- Same feature map size as input size. Do not downsize.
- Lidar pseudo-image are more sparse than natural images.
- Larger kernel and more non-zero pixels than Gaussian. Fill in all pixels inside bbox to a small value. –> CenterPoint enlarged gaussian kernel as well.
- More pixels contribute to offset regression.
- Offset branch not only corrects quantization error, but can also correct regression error. 5x5 pixels to regress offset, as compared to only 1 pixel regressing offset in CenterNet.
- Data augmentation:
- Create a bank of objects
- randomly select 15 GT samples for car/vehicle and place them into the current point cloud.
- Each object is rotated [-9, 9] deg
- global rotation [-45, 45] deg
- Bag of tricks
- High resolution input matters
- High resolution feature map helps
- AdamW + 1 cycle policy for super convergence.
- Data aug during training
- Tricks for winning solution (not in this paper)
- Densification (with pervious 4 frames)
- pointpainting (2D bbox painting and segmentation painting). 50% painted as waymo does not have a rear view cam
- train 3 models and perform ensemble and TTA
- Merging of TTA and ensemble bbox with weighted bbox fusion
Technical details
- 8 bits for orientation (2 bins x (cos + sin + 2-bit cls) per bin) –> This is same as CenterNet.
- Waymo and KITTI uses the same 0.7 3D IoU as the KPI.
Notes