Official implementation of the "PixOOD: Pixel-Level Out-of-Distribution Detection" ECCV 2024 paper

[Paper]

If you use this work please cite:

@InProceedings{Vojir_2024_ECCV,
    author    = {Vojíř, Tomáš and Šochman, Jan and Matas, Jiří},
    title     = {{PixOOD: Pixel-Level Out-of-Distribution Detection}},
    booktitle = {ECCV},
    year      = {2024},
}

• 2024.07.31 - 💥 Code published, inference tested "on my PC and it works" :)
• 2024.07.03 - Accepted to ECCV 2024

• Add documentation and comments
• Clean up and test the training code
• Streamline the training procedure

Download the pre-trained checkpoints (it uses gdown python app)

./checkpoints/download.sh

For example of single frame inference see the example.ipynb

For evaluation of whole sequence see example below:

This command generated the video above (Note that it requires ffmpeg for video generation, otherwise the individual images are stored)

python plot.py --img_dir ./assets/test_seq --out_dir ./_out/vis --dname wos_seq1 --fps 10 --thr 0.995

The training code should work, but I did not test it after refactoring.

The PixOOD first train backbone with MLP classifier for in-distribution classes using configuration stored in ./code/config/dinov2_vit_l.yaml:

cd code
CUDA_VISIBLE_DEVICES=0 python3 train.py --config ./config/dinov2_vit_l.yaml EXPERIMENT.NAME backbone_lp

Then run the Condensation alg. and estimate the decision strategy:

cd code
CUDA_VISIBLE_DEVICES=0 python3 train.py --config ./config/pixood.yaml EXPERIMENT.NAME pixood

Copy the new checkpoints:

cp ./_out/experiments/backbone_lp/checkpoints/checkpoint-best.pth ./checkpoints/checkpoint-backbone.pth
cp ./_out/experiments/pixood/checkpoints/checkpoint-latest.pth ./checkpoints/checkpoint-latest.pth

and the inference code should work. Note that if you change some configuration you will probably need to copy the parameters.yaml from ./_out/experiments/pixood/ to the git repo root directory and change the OUT_DIR to ./.

Copyright (c) 2024 Toyota Motor Europe
Patent Pending. All rights reserved.

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License