RN-50 weights

[seamie6]

Hello,
Are the weights to be used for feature extraction the base RN-50 pre-trained on ImageNet, or is it the BevDet checkpoint? No download link is given

[SBoeder]

Hi,
thanks for the comment. We use the pre-trained weights from BEVDet4D (https://github.com/HuangJunJie2017/BEVDet). You should be able to download the weights from their repository. I will add a download link here as well asap.

[seamie6]

What is the choice for using pre-trained weights from BevDet? Other models utilise ResNet models pre-trained on ImageNet only for fairness. Is it fair to associate LangOcc's performance for this reason, and have you trained the model using the ImageNet only pre-trained ResNet?

[SBoeder]

As I forked my codebase from BEVDet, I simply used the same base config used in the BEVDet repository for occupancy (see here). However, I am very certain that you can achieve very similar results using the standard ImageNet pretrained weights.

[jxl152]

According to the config files, LangOcc uses bevdet-r50-4d-stereo-cbgs.pth.