当前位置:网站首页>【Paper Reading-3D Detection】Fully Convolutional One-Stage 3D Object Detection on LiDAR Range Images
【Paper Reading-3D Detection】Fully Convolutional One-Stage 3D Object Detection on LiDAR Range Images
2022-06-28 06:01:00 【烤粽子】
URL
https://arxiv.org/pdf/2205.13764.pdf
TL;DR
FCOS-LiDAR: Fully Convolutional One-Stage 3D Object Detection on LiDAR Range Images
本文主要提出了lidar-based 的Range View视角下的3D目标检测方法。
主要亮点:
- range view, 只使用标准卷积;
- 在range view,使用了多帧,并且优化方法能有一个比较好的效果;
- Modality-wise Convolutions:通道重新排列;
- 检测头不共享权重;
Dataset/Algorithm/Model/Experiment Detail
网络主体跟模仿FCOS,(其实跟FCOS3D差不多),很多工作在前处理部分,在怎么得到range view image上。
笛卡尔坐标系投影到球坐标系跟其他做法差别不大。
Multi-round Range View Projection
作者发现多帧稠密化之后,range view下有很多的点碰撞,到最后有90%的点被丢弃,导致稠密化之后抽帧效果跟单帧差别不大。作者采用多次(5次效果最好)投影的方式,得到新的 range view image;
Modality-wise Convolutions![[外链图片转存失败,源站可能有防盗链机制,建议将图片保存下来直接上传(img-HPgjx4Pg-1655108238567)(upload://tkyTfmlAkhefI0SrhpSYzT3M1SF.png)]](/img/cc/9e19f8fcc00e202d778815e371f58b.png)
主要思路是把 [x,y,z] , [r,θ,φ] , [i] , [e] and [t] 这几个类型的通道按相同类型的放一起;实验结果如下:mAP涨点0.6~0.7,NDS涨点不多;
Untied Weights of Detection Heads
这里有一点比较有趣检测头不共享权重,image-based方法共享权重效果比较好;
整体网络:
大致跟FCOS3D挺像,backbone有不少修改比如: VR分辨率低减少下采样,使用空洞卷积,ResNet-50 four stages数量更改 (3,4,6,3) --> (4,4,1,1)等。
Experiment



Nuscenes test上效果比centerpoint好
Thoughts
在bev-based方法火热的情况下,出现了一篇range view的方法并且效果还不错,近期range view的工作确实比较少,上次比较有印象的还是RangeDet。
这篇文章的一个目的也是为了避免使用3D卷积,3D稀疏卷积。还把multi frame用上了并且用好了,感觉这一块还能深挖一下。
然后检测头不共享权重似乎也可以深挖一下。
边栏推荐
- How the third-party libraries in cocoapod reference local header files
- Difficulty calculation of Ethereum classic
- Caused by: com.fasterxml.jackson.databind.exc.InvalidDefinitionException: Cannot construct instance
- Jenkins持续集成1
- 6. graduation design temperature and humidity monitoring system (esp8266 + DHT11 +oled real-time upload temperature and humidity data to the public network server and display the real-time temperature
- What is the e-commerce conversion rate so abstract?
- High quality domestic stereo codec cjc8988, pin to pin replaces wm8988
- Openharmony gnawing paper growth plan -- json-rpc
- @Autowired注解为空的原因
- Filecoin黑客松开发者大赛
猜你喜欢

Yygh-8-appointment registration

联想混合云Lenovo xCloud,新企业IT服务门户

Where is the era bonus for developers?

Apple MDM Bypass 免越狱绕过MDM配置锁 免费

Oracle condition, circular statement

6. graduation design temperature and humidity monitoring system (esp8266 + DHT11 +oled real-time upload temperature and humidity data to the public network server and display the real-time temperature

Caused by: com.fasterxml.jackson.databind.exc.InvalidDefinitionException: Cannot construct instance

基于Kotlin+JetPack实现的MVVM框架的示例

YYGH-BUG-03

Jenkins continuous integration 1
随机推荐
使用SQL select count distinct查询语句统计数据库中某个字段的唯一值总数量
Apple MDM Bypass 免越狱绕过MDM配置锁 免费
Simple handwritten debounce function
lombok @EqualsAndHashCode 注解如何让对象.equals()方法只比较部分属性
bash install.sh ********错误
YYGH-6-微信登录
使用SSM框架,配置多个数据库连接
Data warehouse: detailed explanation of hierarchical design
深度學習19種損失函數
Academic search related papers
自定义 cube-ui 弹出框dialog支持多个且多种类型的input框
Data center: Seven Swords of data governance
Filecoin hacker song developer competition
Mosaic data enhanced mosaic
简单手写debounce函数
预训练模型参数不匹配
Working principle of es9023 audio decoding chip
若依实现下拉框
What is the e-commerce conversion rate so abstract?
Slow content advertising: the long-term principle of brand growth