The performance gap in training VirConv-S
jisoo0-0 opened this issue · comments
Hi, thank you for your wonderful work.
BTW, I need some help reproducing your scores on VirConv-S.
I'm trying to train the model from scratch, and I have followed your README file,but I've got below scores.
Car AP@0.70, 0.70, 0.70:
bbox AP:14.4842, 17.3812, 16.8994
bev AP:10.6406, 13.5483, 13.3025
3d AP:9.8955, 11.4659, 11.3564
aos AP:14.12, 16.98, 16.46
Car AP_R40@0.70, 0.70, 0.70:
bbox AP:6.9601, 15.7824, 11.2729
bev AP:5.8445, 11.0559, 7.3793
3d AP:5.3306, 7.8667, 6.2739
aos AP:6.59, 15.19, 10.80
Car AP@0.70, 0.50, 0.50:
bbox AP:14.4842, 17.3812, 16.8994
bev AP:14.5955, 17.4686, 17.0174
3d AP:14.5579, 17.4490, 16.9860
aos AP:14.12, 16.98, 16.46
Car AP_R40@0.70, 0.50, 0.50:
bbox AP:6.9601, 15.7824, 11.2729
bev AP:7.0439, 15.9431, 11.3918
3d AP:7.0190, 15.9104, 11.3695
aos AP:6.59, 15.19, 10.80
I don't know why. Can you please leave me some comments?
I have set the dataset several times and made virtual points with the model you have attached to the README file.
Thanks for reading.
By the way, the results from the VirConv-S model you provided are okay.
I hadsome dataset problems.
The code has no problems at all. Thanks a lot :)