search

noahzn / Lite-Mono

[CVPR2023] Lite-Mono: A Lightweight CNN and Transformer Architecture for Self-Supervised Monocular Depth Estimation
MIT License
540 stars 61 forks source link

Using NYUv2 dataset to train #56

Closed QZZZZZYF closed 1 year ago

QZZZZZYF commented 1 year ago

If I want to use the NYUv2 dataset for training, how do I need to modify the code?

noahzn commented 1 year ago

Hi, thank you for your interest in our work.

My colleague has tried to use my code to train on NYUv2, and here are some of my suggestions:

  1. In the dataset folder you need to prepare your own code to read the data. You need to replace the K (code) with NYUv2's camera intrinsic.
  2. The self-supervised method requires a lot of sequential images for training. You can first pre-train your model using some synthetic indoor datasets.
  3. You might also need to modify the --frame_ids (code) to make sure the camera isn't moving too fast or too slow.
QZZZZZYF commented 1 year ago

Hello，should I prepare NYUDepthV2 datasets following GLPDepth and BTS？ And dataset directory should be: │nyu_depth_v2/ ├──official_splits/ │ ├── test │ ├── train ├──sync/

noahzn commented 1 year ago

Hi, you can arrange the folder structures as you like. I don't know GLPDepth or BTS. As long as you can load the correct images (and ground-truth if you do supervised training), it should work.

QZZZZZYF commented 1 year ago

Hello, I encountered an issue with being unable to read data while trying to train using the NYUv2 dataset. Can your colleague share the code if they succeed?

noahzn commented 1 year ago

Hi, unfortunately, he is not going to share the code. It's his project, and I was not a co-author of that project.