search

pangsu0613 / CLOCs

CLOCs: Camera-LiDAR Object Candidates Fusion for 3D Object Detection
MIT License
351 stars 68 forks source link

Pedestrian val: Bbox, BEV,3D all are 0.0 #63

Closed Rorywh closed 2 years ago

Rorywh commented 2 years ago

Hello, I tried to use the pretrained "mscnn_ped_cyc_trainval_sigmoid_data_scale_1000" and "second_pedestrian_model" to train the model. I have followed the instructions to change some codes in the pedestrian.fhd.config, train.py and the voxelnet.py. The training process is normal except the eval results:

now it is 74150 steps and the cls_loss is : tensor(0.0982, device='cuda:0', grad_fn=) learning_rate: 6.089296127003236e-08 now it is 74200 steps and the cls_loss is : tensor(0.0950, device='cuda:0', grad_fn=) learning_rate: 3.6271128087118313e-08 #################################

EVAL

################################# Generate output labels... validation_loss: 0.008988450088923213 generate label finished(16.10/s). start eval: Pedestrian AP@0.50, 0.50, 0.50: bbox AP:0.00, 0.00, 0.00 bev AP:0.00, 0.00, 0.00 3d AP:0.00, 0.00, 0.00 Pedestrian AP@0.50, 0.25, 0.25: bbox AP:0.00, 0.00, 0.00 bev AP:0.00, 0.00, 0.00 3d AP:0.00, 0.00, 0.00

So, can you help me with this? Thanks for your attention.

Rorywh commented 2 years ago

I have found the reason, and it can output the val results. But the pedestrian 3D AP is lower than expected.

pangsu0613 commented 2 years ago

Hello @Rorywh, thank you for your interests in CLOCs, if the evaluation results are all 0.0, it could be the models are not correctly loaded, could you show more information from the terminal output?

Rorywh commented 2 years ago

Hello @Rorywh, thank you for your interests in CLOCs, if the evaluation results are all 0.0, it could be the models are not correctly loaded, could you show more information from the terminal output?

Thanks for your help. Now the model can output the val results. But it seems that the val results is lower than expected when training and testing on pedestrian.