Giters

jundanl / CGIntrinsics

This is the CGIntrinsics implementation described in the paper "CGIntrinsics: Better Intrinsic Image Decomposition through Physically-Based Rendering, Z. Li and N. Snavely, ECCV 2018".

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

CGIntrinsics

This is the CGIntrinsics implementation described in the paper "CGIntrinsics: Better Intrinsic Image Decomposition through Physically-Based Rendering, Z. Li and N. Snavely, ECCV 2018" (Still Updating, please stay tuned).

The code skeleton is based on "https://github.com/junyanz/pytorch-CycleGAN-and-pix2pix" and "https://github.com/lixx2938/unsupervised-learning-intrinsic-images". If you use our code for academic purposes, please consider citing:

@inproceedings{li2018cgintrinsics,
  	title={CGIntrinsics: Better Intrinsic Image Decomposition through Physically-Based Rendering},
  	author={Zhengqi Li and Noah Snavely},
  	booktitle={European Conference on Computer Vision (ECCV)},
  	year={2018}
}

Dependencies & Compilation:

  • The code was written in Pytorch 0.2 and Python 2, but it should be easy to adapt it to Python 3 version and Pytorch 0.3/0.4 if needed.

Training on the CGIntrinsics dataset:

    python train.py

UPDATES: EASY WAY to get predictions/evaluations on the IIW/SAW test sets:

Since it seems that some people have difficulty running evaluation, we provide precomputed predictions on IIW test set and SAW test set.

    python compute_iiw_whdr.py

(you might need to change judgement_path in this python script to fit to your IIW data path)

    python compute_saw_ap.py

You need modify 'full_root' in this script and to point to the SAW directory you download. To evlaute on unweighted AP% described in the paper, set 'mode = 0' in compute_saw_ap.py and to evaluate on weighted (chanllenging) AP% described in the paper, set 'mode=1' in compute_saw_ap.py.

  • Note: our released model was trained on public released CGI dataset, and if you run on SAW test set, you will get 99.11% for unweighted AP% and 97.93% for weighted AP%, which is slightly better than what was described in the original ECCV camera ready paper. We have updated paper in ArXiv to reflect this change.

Evaluation on the SAW test set:

    python test_saw.py

Note that we only compute AP% (challenge) descirbed in the paper. If you want to compute original AP%, please refer to https://github.com/lixx2938/unsupervised-learning-intrinsic-images

About

This is the CGIntrinsics implementation described in the paper "CGIntrinsics: Better Intrinsic Image Decomposition through Physically-Based Rendering, Z. Li and N. Snavely, ECCV 2018".

Languages

Language:Python 100.0%