Applying and Exploring Supervised Contrastive Regression for Vision-Based Steering Prediction

Self-supervised contrastive learning focus more on classification tasks since the samples from different classes are easy to be distinguished based on the similarity calculation of contrastive loss functions. However, because the continuity nature of regression tasks where samples and labels don't have their associ-ated discreet classes, self-supervised contrastive learning methods face challenges to calculate the similarity between samples for regression tasks. Steering prediction (i.e., a regression task) is a crucial module for self-driving vehicle and U AV autonomous navigation, which is responsible to control the movement orientation. This paper applied the Supervised Contrastive Re-gression (SupCR) method for the vision-based steering prediction. To the best of our knowledge, the SupCR has not been explored in the areas of outdoor self-driving and U AV indoor navigation. The SupCR also obtains better regression performances compared to the bench-mark supervised learning model and other contrastive learning methods. The experiments explore the reasons for better performances of the SupCR according to the visualizations of representation distributions and attention heatmaps.