【Paper Reading-3D Detection】Fully Convolutional One-Stage 3D Object Detection on LiDAR Range Images

URL

https://arxiv.org/pdf/2205.13764.pdf

TL;DR

FCOS-LiDAR: Fully Convolutional One-Stage 3D Object Detection on LiDAR Range Images
This paper mainly puts forward lidar-based Of Range View From the perspective of 3D Target detection method .
Main highlights ：

• range view, Use only standard convolution ;
• stay range view, Used multiple frames , And the optimization method can have a better effect ;
• Modality-wise Convolutions： Channel rearrangement ;
• Detection heads do not share weights ;

Dataset/Algorithm/Model/Experiment Detail

Network subject and imitation FCOS,( In fact, with FCOS3D almost ), A lot of work is in the pre-processing part , In how to get range view image On .
Projection from Cartesian coordinate system to spherical coordinate system is not different from other methods .
Multi-round Range View Projection
The author found that after the multi frame densification ,range view There are a lot of point collisions , In the end 90% Points of are discarded , As a result, the effect of frame extraction after densification is not different from that of a single frame . The author uses many （5 The effect is the best ） The way of projection , Get new range view image;

Modality-wise Convolutions

The main idea is to [x,y,z] , [r,θ,φ] , [i] , [e] and [t] These types of channels are put together in the same type ; The experimental results are as follows ：mAP Rising point 0.6~0.7,NDS Not many rising points ;

Untied Weights of Detection Heads
It is interesting that the detection heads do not share the weight ,image-based The method of sharing weights has a better effect ;

The whole network ：

Roughly the same as FCOS3D It's like ,backbone There are many modifications, such as : VR Low resolution reduces down sampling , Using hole convolution ,ResNet-50 four stages Quantity change (3,4,6,3) --> (4,4,1,1) etc. .

Experiment

Nuscenes test The effect is better than centerpoint good

Thoughts

stay bev-based When the method is hot , There's an article range view And the effect is good , In the near future range view There is really less work , Last time I was more impressed RangeDet.
One of the purposes of this article is to avoid using 3D Convolution ,3D Sparse convolution . Also put multi frame Used and used well , I feel that this piece can be deeply dug .
Then it seems that the detection head can dig deep without sharing the weight .