Supervised Spatially Contrastive Learning

Abstract

Networks pre-trained by supervised contrastive learning have shown high recognition performance in object detection and semantic segmentation. However, existing supervised contrast learning is pre-trained on tasks that capture global features, which may lead to performance bottlenecks due to the gap with downstream tasks where local features are important. Pre-training with unsupervised contrastive learning achieved to improve the accuracy of object detection and semantic segmentation by simply incorporating local features. Therefore, in this paper, we propose a supervised spatially contrastive learning (SSCL). Our proposed method is a modification of an existing supervised contrastive learning to incorporate local features. In our experiments, we compare the proposed method with existing supervised contrastive learning and supervised pre-training using local features in terms of object detection and semantic segmentation accuracy. As a result, in all settings, the network pre-trained by our proposed method outperforms the existing pre-training methods in both downstream tasks.