monocon_rp (2022.08.10)

此项目用于复现monocon 源码地址：https://github.com/chenzihao008/monocon-pytorch.git

修改部分

将训练源码修改为DDP训练,使用mp.spwan启动，单机单卡、单机多卡均可使用

• [修改了以下文件：]
• engine\base_engine.py
• engine\monocon_engine.py
• train.py

存在dataloader多进程读取数据占用显存问题

environment

• [显卡] 3080*1 10G
• [CUDA] 10.1

复现数据集

• [kitti] https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d
• [训练:验证数据比例] 3712:3769

setting

• [batchsize] 8
• [epoch] 200

traget创建方式

训练记录

• [time/epoch] 5-7mins
• [epoch:115] ----------- Eval Results ------------

Pedestrian AP40@0.50, 0.50, 0.50: bbox AP40:49.6264, 39.1815, 32.5803 bev AP40:2.4284, 2.0529, 1.8834 3d AP40:2.1300, 1.7796, 1.3170 aos AP40:39.36, 30.62, 25.27

Pedestrian AP40@0.50, 0.25, 0.25: bbox AP40:49.6264, 39.1815, 32.5803 bev AP40:13.6819, 10.8073, 9.4292 3d AP40:13.1637, 10.4792, 9.0917 aos AP40:39.36, 30.62, 25.27

Cyclist AP40@0.50, 0.50, 0.50: bbox AP40:61.2945, 33.7005, 31.3321 bev AP40:4.5178, 2.2949, 2.1737 3d AP40:3.7567, 1.9922, 1.5229 aos AP40:55.24, 30.34, 28.10

Cyclist AP40@0.50, 0.25, 0.25: bbox AP40:61.2945, 33.7005, 31.3321 bev AP40:19.9975, 10.6099, 9.3342 3d AP40:19.6833, 10.3631, 9.1433 aos AP40:55.24, 30.34, 28.10

Car AP40@0.70, 0.70, 0.70: bbox AP40:96.0738, 78.1678, 70.8113 bev AP40:17.5216, 10.9126, 9.0341 3d AP40:9.7757, 6.3920, 5.1801 aos AP40:94.61, 76.57, 68.84

Car AP40@0.70, 0.50, 0.50: bbox AP40:96.0738, 78.1678, 70.8113 bev AP40:51.2260, 32.1783, 27.4159 3d AP40:44.6220, 28.2990, 22.9361 aos AP40:94.61, 76.57, 68.84

Overall AP40@easy, moderate, hard: bbox AP40:68.9982, 50.3499, 44.9079 bev AP40:8.1559, 5.0868, 4.3637 3d AP40:5.2208, 3.3879, 2.6733 aos AP40:63.07, 45.85, 40.74

• [training_loss]

优化策略及数据对比

知识补充