Skip to content

CityU-AIM-Group/Anna

Repository files navigation

[Paper] [Video Presentation]

By Wuyang Li

Quick Summary

The main idea comes from open-set object detection, where the novel objects are hidden in the background. In OSDA, we do not separate objects from the background since both are out-of-base-class distributions and can be treated as unknown.

  • Even though the source domain only contains the base-class image, we discover novel-class regions hidden in the image to generate unknown signals. This enables unbiased learning in the source domain.
  • With a fine-grained perspective, each image can be treated as the base-class and novel-class regions, regardless of the image-level label. Hence, we align the base and novel class distribution, enabling an unbiased domain transfer.
  • We use the causal theory to guide the method design.

Experimental Environment

  • cudatoolkit == 10.0
  • torch == 1.6
  • torchvision == 0.7.0
  • numpy == 1.21.4
  • scikit-learn == 1.0.2

Get Start

  • Download the Officehome dataset.
  • Change the data root in train.py: --data-root
  • Run run.sh for all sub-tasks.
  • Generate final results in latex format.

Reproduced resuts by us:

A $\rightarrow$ C A $\rightarrow$ P A $\rightarrow$ R C $\rightarrow$ A C $\rightarrow$ P C $\rightarrow$ R P $\rightarrow$ A P $\rightarrow$ C P $\rightarrow$ R R $\rightarrow$ A R $\rightarrow$ C R $\rightarrow$ P Avg
69.3 73.2 76.3 64.7 68.6 72.7 65.9 63.9 76.0 70.6 68.1 78.7 70.7
  • The performance of each sub-task is slightly different from the paper due to different seeds, environments, and warm-up iterations.

  • The average performance is the same.

Limitations and Disussions

  • The hyperparameters, e.g., top-K and gradient scalers, are relatively sensitive to the dataset properties (especially for tiny datasets with homogeneous scenes) and warm-up stages.
  • We tried to avoid introducing extra parameters in the inference stage, which is sub-optimal. Using a different classification head and introduce other designs for the unknown prediction will be better.
  • The idea of dicovering unknown components in a base-class image can be transferred to other tasks.

Contact

If you have any questions or ideas you would like to discuss with me, feel free to let me know through wuyangli2-c @ my.cityu.edu.hk. Except for the main experiment on Officehome, other tiny-scaled benchmark settings will be released later if needed.

Citation

If this work is helpful for your project, please give it a star and citation. Thanks~

@InProceedings{Li_2023_CVPR,
    author    = {Li, Wuyang and Liu, Jie and Han, Bo and Yuan, Yixuan},
    title     = {Adjustment and Alignment for Unbiased Open Set Domain Adaptation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2023},
    pages     = {24110-24119}
}

Abstract

Open Set Domain Adaptation (OSDA) transfers the model from a label-rich domain to a label-free one containing novel-class samples. Existing OSDA works overlook abundant novel-class semantics hidden in the source domain, leading to a biased model learning and transfer. Although the causality has been studied to remove the semantic-level bias, the non-available novel-class samples result in the failure of existing causal solutions in OSDA. To break through this barrier, we propose a novel causalitydriven solution with the unexplored front-door adjustment theory, and then implement it with a theoretically grounded framework, coined Adjustment and Alignment (ANNA), to achieve an unbiased OSDA. In a nutshell, ANNA consists of Front-Door Adjustment (FDA) to correct the biased learning in the source domain and Decoupled Causal Alignment (DCA) to transfer the model unbiasedly. On the one hand, FDA delves into fine-grained visual blocks to discover novel-class regions hidden in the base-class image. Then, it corrects the biased model optimization by implementing causal debiasing. On the other hand, DCA disentangles the base-class and novel-class regions with orthogonal masks, and then adapts the decoupled distribution for an unbiased model transfer. Extensive experiments show that ANNA achieves state-of-the-art results.

image

About

[CVPR' 23] Adjustment and Alignment for Unbiased Open Set Domain Adaptation

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published