June 2019
tl;dr: Use auxiliary tasks (ground estimation and depth completion) and sensor fusion boost 3D object detection. MMF is fast running at 13 FPS.
Overall impression
This paper is built on ContFuse and two-stage sensor fusion methods such as MV3D and AVOD. MMF and ContFuse is similar to AVOD that it uses fused feature for proposal generation. And MMF and ContFuse method is anchor-free. However MMF is better than ContFuse in that it uses depth estimation for a dense pseudo-lidar point cloud.
The paper is also influenced by HDNet which exploits HDmap and estimates ground height for 3D detection.
MV3D –> AVOD –> ContFuse –> MMF
This boost the 2D hard by more than 8 AP (from 80 to 88) among real-time models. (RRC from sensetime performs really well for 2D OD but runs at 3.6 s/frame)
Key ideas
- Cascaded design (such as F-Pointnet) suffers from the drawback from each stage. It has difficulty recovering object that are far away.
- Sensor fusion architecture: point-wise and ROI-wise feature fusion.
- pointwise sensor fusion: use lidar points to establish correspondence between BEV and image (like in contFuse). Multi-scale features from FPN of image is fused into each layer of lidar feat maps.
- ROI-wise sensor fusion: project 3D bbox to BEV and image space (with RoIAlign and oriented RoIAlign)
- Proposal generation: 1x1 conv on top of last BEV feat map (like YOLO). This feat map has feat from both image and BEV.
- Map fusion: estimated ground height is subtracted from the regression target. This eases the 3D object localization (don’t we assume this from the beginning?)
- The image backbone is small (Resnet 18) but it benefits quite a bit from lidar sensor fusion.
Technical details
- The depth completion idea is very similar to pseudo-lidar. A depth map GT is used to supervise this task.
- By applying ground estimation and depth completion during training, the network learns a better representation, but no need to deploy during inference.
- Pretraining: resnet-18 and pretrained depth estimation
Notes
- 3D perception matters more than 2D as motion planning happens in bird’s eye view.
- Honestly speaking, without the bbox finetuning, the performance is already pretty good for industry applications. This way we could harvest the benefit of Sensor fusion more easily.
- Q: why it is so fast? –> ResNet18 as backbone
- Q: how to do point wise feature fusion without accurate calibration?