COMMITTEE CHAIR: Dr. Suxia Cui
CO-COMMITTEE CHAIR: Dr. Lujun Zhai
TITLE: EFFECTIVE DATA AUGMENTATION STRATEGIES FOR SMALL OBJECT DETECTION
ABSTRACT: Data augmentation remains the most effective method of improving object detection, particularly in scenes where small objects dominate and the annotated datasets are small. While geometric augmentation techniques such as flipping, cropping, and photometric augmentations such as brightness, color jittering adjustments, and content-level augmentations such as mixup and copy-paste augmentation have proved useful, current research suggests the viability of learned augmentation policies, instance-level generative augmentations, and diffusion-based augmentations. Here, we present a structured comparison of the augmentation methods on the YOLO11m detector, focusing particularly on the detection of small objects on the SODA-D dataset. We benchmark the standard YOLO11m model with models learned with (i) conventional transforms, (ii) learned augmentation policies, (iii) diffusion-based semantic editing, and (iv) instance-level redrawing of images by employing pre-trained generative models. Through different experimental configurations, we see that dataset augmentation by instance-level augmentations always outperforms the rest, achieving the highest boosts in mean Average Precision (mAP) for object categories of small size. Our research also indicates that augmentations at the object-instance level guarantee preservation of scene context and introduce visual diversity, which overcomes overfitting and improves robustness. These findings provide empirical evidence for instance-aware generative augmentations being a robust and generalizable method of detecting small objects, and leave the possibility of data-efficient training pipelines for future vision tasks.
Keywords: Small Object Detection, Data Augmentation, Generative Models, YOLO11m
Room Location: Electrical Engineering Conference Room 315D