Rethinking Text-to-Image as Semantic-Aware Data Augmentation for Indoor Scene Recognition

Researchers have proposed novel approaches to improve image recognition and data augmentation using Stable Diffusion and modified geometric transformation functions.

A team of researchers introduced a novel approach leveraging Stable Diffusion (SD) to generate synthetic images for indoor scene recognition, addressing challenges posed by lighting conditions, occlusions, and diverse object arrangements^[1]. The proposed method enhances the training of deep models for indoor image recognition by enriching the training data pool with diverse and realistic synthetic indoor scenes. To prevent the misuse of SD synthetic images, a counter measure based on Diffusion Reconstruction Error (DIRE) was introduced, enabling the training of robust classifiers using lightweight deep models. Experimental findings on the MIT Indoor Scene dataset showed that the approach can perfectly recognize SD-generated images with 100% accuracy using MobilenetV3^[1]. In a related development, another research team proposed a modified geometric transformation function to replace nearest neighbor interpolation in data augmentation, addressing risks of pixel-level annotation errors and low-pass filtering effects^[2]. The authors also implemented an offline data augmentation pipeline to generate interpolation-specific augmented training data and integrated a mean-based class filtering mechanism to handle undefined categorical labels. Experimental evaluation on three medical image segmentation datasets and the XBAT+ datasets demonstrated performance gains across multiple quantitative metrics^[2].