Guiding Pseudo labels with Uncertainty Estimation for Source-free Unsupervised Domain Adaptation (CVPR 2023)

This is the official implementation of the CVPR 2023 paper "Guiding Pseudo-labels with Uncertainty Estimation for Source-free Unsupervised Domain Adaptation" - Mattia Litrico, Alessio Del Bue, Pietro Morerio.

https://arxiv.org/abs/2303.03770

@inproceedings{litrico_2023_CVPR,
  title={Guiding Pseudo-labels with Uncertainty Estimation for Source-free Unsupervised Domain Adaptation},
  author={Litrico, Mattia and Del Bue, Alessio and Morerio, Pietro},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2023}
}

Abstract

Standard Unsupervised Domain Adaptation (UDA) methods assume the availability of both source and target data during the adaptation. In this work, we investigate Source-free Unsupervised Domain Adaptation (SF-UDA), a specific case of UDA where a model is adapted to a target domain without access to source data. We propose a novel approach for the SF-UDA setting based on a loss reweighting strategy that brings robustness against the noise that inevitably affects the pseudo-labels. The classification loss is reweighted based on the reliability of the pseudo-labels that is measured by estimating their uncertainty. Guided by such reweighting strategy, the pseudo-labels are progressively refined by aggregating knowledge from neighbouring samples. Furthermore, a self-supervised contrastive framework is leveraged as a target space regulariser to enhance such knowledge aggregation. A novel negative pairs exclusion strategy is proposed to identify and exclude negative pairs made of samples sharing the same class, even in presence of some noise in the pseudo-labels. Our method outperforms previous methods on three major benchmarks by a large margin. We set the new SF-UDA state-of-the-art on VisDA-C and DomainNet with a performance gain of +1.8% on both benchmarks and on PACS with +12.3% in the single-source setting and +6.6% in\ multi-target adaptation. Additional analyses demonstrate that the proposed approach is robust to the noise, which results in significantly more accurate pseudo-labels compared to state-of-the-art approaches.

Data Preparation

1. Please download the VisDA-C and DomainNet dataset, and put it under ${DATA_ROOT}. By default ${DATA_ROOT} is set to data. The prepared directory would look like:

${DATA_ROOT}
├── VISDA-C
│   ├── train
│   ├── validation
│   ├── test

2. Please download the DomainNet dataset, and put it under ${DATA_ROOT}. Notice that we follow MME to use a subset that contains 126 classes from 4 domains, so we also compiled .txt files for your convenience based on the the image labels, provided under ./datasets/domainnet-126/ The prepared directory would look like:

${DATA_ROOT}
├── domainnet-126
│   ├── real
│   ├── sketch
│   ├── clipart
│   ├── painting
│   ├── real_list.txt
│   ├── sketch_list.txt
│   ├── clipart_list.txt
│   ├── painting_list.txt

VisDA-C

Training

VISDA-C experiments are done for train to validation adaptation. Before the test-time adaptation, we should have the source model. You may train the source model with script train_VISDA-C_source.sh as shown below. After obtaining the source models, you will find model weights inside the directory logs. Now run train_VISDA-C_target.sh to execute test-time adaptation.

# train source model
bash train_VISDA-C_source.sh

# train TTA
bash train_VISDA-C_target.sh

If you want to change the default ${DATA_ROOT}, please use the following:

# train source model
bash train_VISDA-C_source.sh <DATA_ROOT>

# train TTA
bash train_VISDA-C_target.sh <DATA_ROOT>

This will reproduce Table. 3 from the main paper:

DomainNet-126

Training

DomainNet-126 experiments are done for 7 domain shifts constructed from combinations of Real, Sketch, Clipart, and Painting. Before the test-time adaptation, we should have the source model. You may train the source model with script train_domainnet-126_source.sh as shown below. After obtaining the source models, you will find model weights inside the directory logs. Now run train_domainnet-126_target.sh to execute test-time adaptation.

# train source model
# example: bash train_VISDA-C_source.sh real
bash train_domainnet-126_source.sh <SOURCE_DOMAIN>

# train SF-UDA
# example: bash train_VISDA-C_target.sh real sketch
bash train_domainnet-126_target.sh <SOURCE_DOMAIN> <TARGET_DOMAIN>

If you want to change the default ${DATA_ROOT}, please use the following:

# train source model
# example: bash train_VISDA-C_source.sh real
bash train_domainnet-126_source.sh <SOURCE_DOMAIN> <DATA_ROOT>

# train SF-UDA
# example: bash train_VISDA-C_target.sh real sketch
bash train_domainnet-126_target.sh <SOURCE_DOMAIN> <TARGET_DOMAIN> <DATA_ROOT>

This will reproduce Table. 4 from the main paper:

License

GNU GENERAL PUBLIC LICENSE Version 3, 29 June 2007 Copyright © 2007 Free Software Foundation, Inc. http://fsf.org/