Skip to content

Baidu Rope3d detector based on yolov7

License

Notifications You must be signed in to change notification settings

zqs1996/yolov7_3d

 
 

Repository files navigation

基于YOLOv7的单目3D检测方案

本方案尚未完成,当前工作: 3d指标正确评测

一、数据集

百度rope3d,路端数据

二、模型架构

先采用基础版的YOLOv7,后续可升级YOLOv7 E6E

三、损失函数

除原有的2d损失(cls, obj, bbox)之外,增加3dbbox损失

代码工作:

  1. build_targets

    根据loss选择,进行标签转换

  2. 3d bbox loss

    3d中心点投影偏移:L1

    深度:LaplacianAleatoricUncertaintyLoss

    3d尺寸:L1(DimAwareL1经常出现梯度问题)

    角度:L1

    后续有待优化

  3. 损失权重,需进行实验设置

四、训练过程

按照常规端到端训练方法进行训练和调参

五、初步效果

标签gt

pred

About

Baidu Rope3d detector based on yolov7

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Jupyter Notebook 98.4%
  • Python 1.6%