fabiotosi92 / NeRF-Supervised-Deep-Stereo

A novel paradigm for collecting and generating stereo training data using neural rendering
https://nerfstereo.github.io/
MIT License
348 stars 19 forks source link

Request tips for model training #14

Closed chewry closed 1 year ago

chewry commented 1 year ago

Thank you for sharing of your nice work!

Inspired by your work, I finetuned the model to apply this to another domain. Unfortunately, fine-tuning failed. In order to check whether it is a domain problem, we fine-tuned the model on the provided NeRF-stereo triplet dataset, but failed as well. (Detection of texture rather than object boundary)

Because there is no code for the training, I used the same hyperparameter and augmentation procedures as RAFT-stereo, as written in your paper. Is there anything else to note for training? If you have any tips for training, please give me some advice.

fabiotosi92 commented 1 year ago

Hi @chewry, could you please be more specific? What do you mean by fine-tuning failing? Does the network forget how to perform stereo matching or does it simply worsen the disparity estimates?

chewry commented 1 year ago

Yes. the network forget how to perform stereo matching. The finetuned model detects texture of the objects. I use the same augmentation procedures as RAFT-stereo except _resize_sparse_flowmap. (The sparse resizing makes the results worse. Textures regions are more emphasized.)

Did you use the all augmentations in RAFT-stereo (color transform, erasing regions, resizing, horizontal flip, ...)?

mattpoggi commented 1 year ago

Hi, Can you share an output from your finetuned model?