OSA: Object-level Scale Alignment for Small Object Detection in Large-Scale Images
Yuxiang Wang, Y. Y. Ji, X. R. Chen, Chuanyuan Tan, Li Yajin, Haozhong Xue, Zheng Zhao
Abstract
Detecting small objects (e.g., those smaller than 20 × 20 pixels) in large-scale images remains a significant and challenging problem. Modern CNN-based detectors often struggle due to the scale mismatch between pre-training datasets and the target dataset used for detector fine-tuning. In this paper, we investigate the impact of scale alignment between pre-training and target datasets, and introduce an effective Object scale alignmenting (OSA) approach to align object scales across datasets, thereby enhancing the representation of small objects. Furthermore, we propose a refined version of the OSA method, termed OSA-RF, which improves scale alignment from the image level to the instance level, strengthening the similarity between pre-training and target datasets. To address the potential distortion of image structure caused by traditional scale alignmenting, we introduce a novel Probabilistic Structure Inpainting (PSI) method for background processing. Extensive experiments across multiple detectors demonstrate that OSA-RF significantly improves performance on the TinyPerson dataset and outperforms state-of-the-art detectors by a considerable margin.