Fangchang Ma

> Can you give us an example? Akash gave an excellent example. Also, please do not post the same question repeatedly.

Hi. Hopefully the data directory structure is clear from the README file (copied and pasted here): ``` . ├── self-supervised-depth-completion ├── data | ├── kitti_depth | | ├── train |...

> what should be the structure of the data folder in that case The data folder structure is illustrated in the [README](https://github.com/fangchangma/self-supervised-depth-completion#requirements). You can switch to any data loader, as...

Thanks @guivenca ! The MR looks great to me, except for two minor comments: 1. The name "reduce" is a bit vague - can we change it to something along...

You can refer to our earlier paper on [Sparse-to-Dense: Depth Prediction from Sparse Depth Samples and a Single Image](https://arxiv.org/pdf/1709.07492.pdf) (e.g., Fig 5) for such comparisons.

For cameras with different intrinsic parameters than the KITTI cameras, there is a slightly more complicated image resizing process than simple cropping. In particular, we need to resize the image...

@XiaotaoGuo It might or might not work well with a different set of intrinsics, but there is simply no guarantee that the trained network would transfer directly to this new...

> What if we use our own dataset to train the network and test with it? Then there is no need for any resizing

Hi. - We used the entire training dataset. - We used 352 x 1216, but downsampling to half or even smaller sizes should also work fine. - Simple data augmentation...

This would require minor changes to `conditional_save_pred(self, mode, i, pred, epoch)` [[link](https://github.com/fangchangma/self-supervised-depth-completion/blob/48a73af3b87e35b7c33d11b1659a57ac82d43a0a/helper.py#L170)]