Quality and Details For Depth Reconstruction #1
Description
Hello, thank you so much for making this great work open source.
I encountered some minor issues while reproducing this work. I downloaded the Downtown (anymal) dataset from the TartanAirV2 dataset and used your train_single script to train the depth encoder and decoder with the pretrain.yaml.
During training, I observed that the training_loss dropped to around 2 and converged (my epoch is 200). When I opened the saved images, I found that the reconstruction effect was acceptable, but there were still significant problems in the details (I haven't tested it on a real sensor yet).
I would like to ask the author for some suggestions and whether the pretrain in the paper was trained on the entire TartanAirV2 dataset. What were the epoch and other parameters? Also, could you provide the pre-trained pth as mentioned in the readme?
