VAD

VAD copied to clipboard

plan_L2_1s:0.3348474009348846 plan_L2_2s:0.6005098198601633 plan_L2_3s:0.9474086193724374 plan_obj_col_1s:0.0 plan_obj_col_2s:0.0 plan_obj_col_3s:3.255844337443258e-05 plan_obj_box_col_1s:0.0019535065442469234 plan_obj_box_col_2s:0.002979097479976558 plan_obj_box_col_3s:0.006088428835334152

projects/configs/VAD/VAD_base_e2e.py

L2 is better， but collasion is worse than the paper‘s. Why?

loss_plan_col=dict(type='PlanCollisionLoss', loss_weight=1.0),

can I make the above "loss_weight" bigger?

Did you train the model yourself, or use the pretrained model? If using the pretrained one, I think VAD_base_stage_2.py should be used instead of VAD_base_e2e.py Additionally, I think that they use another image normalization config in the paper and pretrained model. See https://github.com/hustvl/VAD/issues/18#issuecomment-1711608585 and https://github.com/hustvl/VAD/issues/9#issuecomment-1696708154

I train the model from resnet50-19c8e357.pth by the configs of "VAD_base_e2e.py“. Is the "loss_weight" not important?

I was in a similar situation

-------------- Motion Prediction -------------- EPA_car: 0.6096603478939834 EPA_pedestrian: 0.3235923685435086 ADE_car: 0.8396922945976257 ADE_pedestrian: 0.8586953282356262 FDE_car: 1.1747483015060425 FDE_pedestrian: 1.189048409461975 MR_car: 0.12827822120866592 MR_pedestrian: 0.16290983606557377

-------------- Planning -------------- gt_car:4.503418636452432 gt_pedestrian:2.099042781793319 cnt_ade_car:3.6200429771439735 cnt_ade_pedestrian:1.077358859152178 cnt_fde_car:3.4264504786091035 cnt_fde_pedestrian:0.9533111935924985 hit_car:2.9869115061535454 hit_pedestrian:0.7980074233248682 fp_car:0.40378980269583903 fp_pedestrian:0.20296932994725533 ADE_car:3.1434736251831055 ADE_pedestrian:0.9509602189064026 FDE_car:4.025217056274414 FDE_pedestrian:1.1335331201553345 MR_car:0.43953897245555773 MR_pedestrian:0.1553037702676304 plan_L2_1s:0.30277927277533684 plan_L2_2s:0.5726086170295752 plan_L2_3s:0.949615497207263 plan_obj_col_1s:0.0 plan_obj_col_2s:0.0 plan_obj_col_3s:3.255844337443258e-05 plan_obj_box_col_1s:0.0034186364524321157 plan_obj_box_col_2s:0.005176792342254347 plan_obj_box_col_3s:0.011851273242745881 fut_valid_flag:1.0

projects//configs/VAD/VAD_tiny_e2e.py

@StevenJ308, I have checked my log file, not find redownload the resnet50 model. Maybe, some superparameters are not same as the paper's.

I would like to ask why I can not use my own training pth file to test? I still get errors when I use my own pth:result_dict['ADE_'+cls] = all_metric_dict['ADE_'+cls] / all_metric_dict['cnt_ade_'+cls] ZeroDivisionError: float division by zero

您是自己训练模型，还是使用预训练模型？如果使用预训练模型，我认为VAD_base_stage_2.py应该使用而不是VAD_base_e2e.py 此外，我认为他们在论文和预训练模型中使用了另一个图像规范化配置。请参阅#18（评论）和#9（评论） Hello, I would like to ask what is the difference between the pth trained by myself and the pth provided by the paper?