SPARC: SuperPixel-Aware Region Contrastive Learning for Self-Supervised Dense Prediction

Summary

SPARC is a self-supervised pretraining method that adds a region-level contrastive branch to the DenseCL/MoCo-v2 framework. Each training image is segmented into SLIC superpixels; region identities are propagated through the geometric augmentations, and backbone features are pooled within each region to form region embeddings. The same superpixel ID across the two augmented views forms a positive pair, all other region embeddings act as negatives, and the region loss is combined with the global InfoNCE objective through a weighting λ. With ResNet-18 pretrained for 100 epochs on COCO or ImageNet100 and transferred to PASCAL VOC, SPARC beats both baselines in every reported configuration: with COCO pretraining and 25 fine-tuning epochs it reaches 33.78 mIoU segmentation (DenseCL 29.57, MoCo-v2 25.81) and 27.67 AP detection (25.48, 24.84). A λ ablation shows pure region-level contrast outperforming pure pixel-level and pure image-level objectives. The limitations are substantial: only ResNet-18 is tested, pretraining corpora top out at 135k images, and transfer is measured on VOC alone; the closest region-level prior work (DetCon) is neither cited nor compared, modern baselines such as DINO and MAE are absent, the only ablated hyperparameter is λ, the 'computationally efficient' claim carries no timing data, and the number of seeds behind the ±std values is unstated.