Do you have the plan to opensource the training code? about fast-depth

[fangchangma]

We might release the training code in the future. For now, please refer to https://github.com/fangchangma/sparse-to-dense.pytorch for training.

[v-raja]

Hey @fangchangma and @dwofk,

Great job on the paper! I tried to understand how to train the network based on the information in the link above, but I didn't understand if I need to provide depth data to train the model. (Sorry if this is obvious, but I've looked through the code and was unable to figure this out.)
Thanks!

[dwofk]

Hi @Vivkr,

If you are referring to providing sparse depth data to train the model, then no, you do not need to provide that. You can use the 'rgb' modality. The sparsifier and number of sparse depth samples become irrelevant.

Models in this project support 'RGB' input, not 'RGBd' input.

[SHRHarry]

Hi @dwofk and @fangchangma ,
I'm interested if I want to train my own dataset, is it possible to use the ground truth collected by ZED camera?
Thanks!

[HuynhLam]

Hi,
I just want to make sure that the detailed training for all encoder and decoder permutations is based on the spare-to-dense paper, isn't it?
To be specific:

• The loss was the berHu loss (similar to Laina et al paper)
• Data augmented ...
• Init using ImageNet-pretrained models
• Initial lr = 0.01
• Train on 20 epochs
  .
  .
  Additionally, this is a great work. Thanks!

[dwofk]

Hi @HarryShen19950808

Model definitions are provided here. If you have your own dataset, you can try modifying the models as necessary and retraining for your particular dataset.

[dwofk]

Hi @HuynhLam

Yes, the training for our models is based on the sparse-to-dense work.

• no, we use L1 loss (note that the cited sparse-to-dense work also uses L1 loss)
• yes, the data augmentation is the same
• yes, encoder weights are initialized using ImageNet-pretrained models
• yes, initial lr = 0.01
• yes, we train for 20 epochs