Domain Generalization via Inference-time Label-Preserving Target Projections

Code for the CVPR 2021 Oral paper "Domain Generalization via Inference-time Label-Preserving Target Projections". The proposed method introduces a new way to handle Domain Generalization problem as compared to the traditional methods. It uses test-time optimization to optimize the target features and projects them in the source manifold.

Authors: Prashant Pandey, Mrigank Raman*, Sumanth Varambally*, Prathosh AP.

(* denotes equal contribution)

Requirements

• Python 3.6.10
• PyTorch version 1.6.0
• CUDA version 10.1
• 4 NVIDIA® Tesla® V100(16 GB Memory) GPUs.

Usage

Train {dataset}_{backbone}_FNet.py using source domains to get domain agnostic representations

python {dataset}_{backbone}_FNet.py

Learn generative model on features from FNet and perform Target projections

python {dataset}_{backbone}_Gphi_projection.py

To do Nearest Neighbor search on VLCS with FNet features

python vlcs_1NN_Sampler.py

Citation

If you find our work useful, please consider citing our paper.

@InProceedings{Pandey_2021_CVPR,
    author    = {Pandey, Prashant and Raman, Mrigank and Varambally, Sumanth and AP, Prathosh},
    title     = {Generalization on Unseen Domains via Inference-Time Label-Preserving Target Projections},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2021},
    pages     = {12924-12933}
}

For clarifications, contact Prashant Pandey