January 2020
tl;dr: Fuse radar to camera with sparse pseudo-image as input and two output branches for small and large object detection.
Overall impression
This paper uses similar method to convert radar pins into pseudo-image as in distant object detection. It is called “sparse radar image”.
Critics: The “dense radar image” does not make sense to me as it warps a 169-dim feature into a 13x13 image and apply 2D conv. There is no guarantee of the order in the 169-d feature and thus 2D conv seems a bit random.
Key ideas
- Input: Image (416x416x3), Radar (416x416x3) (depth, lateral velocity, longitudinal velocity). Both velocities are compensated by ego velocity. Alternatively, radar can be warped into a 13x13 pseudo image directly, which does not make sense. (see above).
- Architecture
- Image features extracted by TinyYolov3
- Radar features extracted by a VGG like structure.
- Feature maps fused at 13x13 (downsize x8) level and has big object detection branch
- Fused feature map upsampled by 2 to 26x26 and has small object detection
- Fusing radar leads to better object detection. However it did not boost individual class detection. Radar signal is not good at identifying classes in a multi-class classification setup.
Technical details
- RVNet uses fixed radar number, 169, which is a bit puzzling. The number of points in the radar data per frame is not fixed. Using this class to load nuscenes radar pcd data, and the point cloud is in format of 18 (attr. num) x
num_points, and num_points varies (63, 47, 48, etc).
- Sparse radar image is better than dense radar image across the board (as expected).
- The one with two branches is better than one detection branch
Notes
- Questions and notes on how to improve/revise the current work