We found that using weights from our lower input resolution (256 x 512) depth model as initialization leads to faster convergence, rather than training from scratch on full reolsution (512 x 1024) which takes significant amount of computational time. Therefore, we can pre-initialize using these weights for training the final model which took only 5 epochs of training. Please download these weights from this google drive link. This directory contain following weight files:
- depth.pth
- encoder.pth
- albedo.pth
- ambient.pth
- pose.pth
- pose_encoder.pth