Reproduce “Test-Time Modification: Inverse Domain Transformation for Robust Perception” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Arpit Jadon, Joshua Niemeijer, Yuki M. Asano
Researchers from German Aerospace Center and University of Technology Nuremberg developed Test-Time Modification (TTM), an inference-time method that employs generative image-to-image models to transform target-domain images back to the source distribution. TTM achieved mIoU improvements of up to 17.7% for semantic segmentation, mAP@50 gains of up to 21.6% for object detection, and Top-1 accuracy increases of up to 24.7% for image classification on various domain shift benchmarks.
Key insights
- Large-scale generative image-to-image models can be effectively employed at inference time to invert domain shifts, mapping diverse target images back to a defined source distribution.
- Improving robustness against domain shift can be achieved by transforming the input data to match the discriminative model's training distribution, rather than modifying the model itself.
- TTM reduces both epistemic and aleatoric uncertainty by enhancing the visual quality of challenging inputs, making them more interpretable for downstream perception tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.