How to Grow Small Vision Datasets with SAM 3 and Synthetic Augmentation

The text discusses a data-centric approach to enhancing small vision datasets by utilizing the Segment Anything Model 3 (SAM 3) and Roboflow to apply segmentation-aware augmentation. It highlights the inadequacy of traditional image-level augmentations, which often fail to introduce necessary variation for model generalization, and presents SAM 3 as a solution for generating high-quality instance segmentation masks that improve label quality. The tutorial demonstrates this by using an Aerial Solar Panels dataset, where SAM 3 is employed to transform coarse object detection labels into refined segmentation masks, which are then validated and cleaned. This process ensures that geometric and photometric augmentations are applied at the object level, preserving label integrity and introducing realistic variation without the introduction of noise. By employing dataset versioning, the tutorial emphasizes the importance of maintaining data integrity and measuring the impact of augmentation on model performance, showcasing how this method can increase recall and detection quality without significantly raising false positives. The approach is particularly beneficial for datasets that are visually distinct, small, or costly to expand, and stresses the importance of quality data and validation over dataset size when resources are limited.