Code for our 2025 paper "Effective Dual-Region Augmentation for Reduced Reliance on Large Amounts of Labeled Data," by Prasanna Reddy Pulakurthi, Majid Rabbani, Celso M. de Melo, Sohail A. Dianat, and Raghuveer Rao. [PDF]
Keywords: Data Augmentation, Classification, Source-Free Domain Adaptation (SFDA), Person Re-Identification (ReID)
Demo - [Hugging Face Spaces]
An interactive demo for generating SPM augmentation.
We introduce a novel dual-region augmentation approach that reduces reliance on large-scale labeled datasets while improving model robustness across diverse computer vision tasks.
Our method applies:
- Random noise perturbations to foreground objects.
- Spatial shuffling of background patches.
This structured augmentation improves model generalization and robustness without requiring additional supervision.
Evaluations on the PACS dataset (SFDA) and Market-1501, DukeMTMC-reID datasets (ReID) show consistent improvements over existing methods across single-target and multi-target settings.
Overall pipeline of the proposed augmentation process: foreground noise + background shuffle.
We apply random patch noise to the foreground and spatial shuffling to the background to create diverse augmented images.
Step-by-step illustration of the Foreground-Background Augmentation method: Input Images → U2Net Mask Extraction → Foreground Patch Noise → Background Patch Shuffle → Final Augmented Output.
Our augmentation method can be applied across a variety of tasks, including Source-Free Domain Adaptation (SFDA) and Person Re-Identification (ReID).
Examples of augmented outputs generated by applying Foreground Patch Noise and Background Patch Shuffling.
Implementation details and training scripts are available in ./SFDA/.
Implementation details and training scripts are available in ./Person_ReID/.
Our method achieves state-of-the-art performance across Source-Free Domain Adaptation and Person ReID benchmarks.
We evaluate on the PACS dataset for both single-target and multi-target domain adaptation.
Classification accuracy (%) on PACS dataset. Our method achieves the highest accuracy across both settings.
We evaluate across ResNet-18 and EfficientNet-b4 backbones, consistently outperforming baselines and existing augmentation techniques.
Person ReID performance on Market-1501 and DukeMTMC-reID datasets. Our augmentation strategy achieves superior results across all metrics.
If you find this work useful, please cite:
@inproceedings{pulakurthi2025effective,
title={Effective dual-region augmentation for reduced reliance on large amounts of labeled data},
author={Pulakurthi, Prasanna Reddy and Rabbani, Majid and De Melo, Celso M and Dianat, Sohail A and Rao, Raghuveer M},
booktitle={Synthetic Data for Artificial Intelligence and Machine Learning: Tools, Techniques, and Applications III},
volume={13459},
pages={210--218},
year={2025},
organization={SPIE},
doi = {10.1117/12.3058627},
URL = {https://doi.org/10.1117/12.3058627}
}