search

TRAILab / CaDDN

Categorical Depth Distribution Network for Monocular 3D Object Detection (CVPR 2021 Oral)
Apache License 2.0
366 stars 62 forks source link

I test the caddn.pth model, the moderate model 3D AP is 15.93, but the site say it's 16.07. #70

Closed rockywind closed 3 years ago

rockywind commented 3 years ago

My Environment is: ubuntu 16.04 cuda11.1 pytorch1.7.1 Per_GPU_batch_size=4 GPU_num=4

2021-09-10 11:25:34,568 INFO
Car AP@0.70, 0.70, 0.70: bbox AP:89.9561, 80.0459, 78.7131 bev AP:34.7719, 25.4463, 24.0320 3d AP:27.8585, 21.2847, 18.5648 aos AP:89.11, 78.92, 76.98 Car AP_R40@0.70, 0.70, 0.70: bbox AP:95.3015, 82.5968, 77.3852 bev AP:31.3691, 21.3940, 19.2927 3d AP:23.5936, 15.9324, 13.4591 aos AP:94.28, 81.29, 75.64 Car AP@0.70, 0.50, 0.50: bbox AP:89.9561, 80.0459, 78.7131 bev AP:62.5091, 46.2527, 44.8870 3d AP:57.8045, 43.3583, 37.7294 aos AP:89.11, 78.92, 76.98 Car AP_R40@0.70, 0.50, 0.50: bbox AP:95.3015, 82.5968, 77.3852 bev AP:62.8995, 46.3063, 42.3390 3d AP:57.0173, 40.6578, 36.8525 aos AP:94.28, 81.29, 75.64

codyreading commented 3 years ago

It's possible that there are small floating point differences due to different environment versions/hardware. I've experienced different results when running the exact same code with different environments/hardware, and have achieved slightly different results. This difference is quite negligible, I wouldn't worry about it.

rockywind commented 3 years ago

The result is that I tested the pre-train model downloaded from GitHub. The setting is: per_GPU_size=12 GPU_num=3 image

codyreading commented 3 years ago

Okay that looks good.