Contrastive Class Anchor Learning for Open Set Object Recognition in Driving Scenes

Abstract

Conventional object recognition models operate under closed-set assumptions presuming that the training dataset is sufficiently comprehensive that any object detected during inference can be assigned to some known prior class. This assumption is flawed and potentially dangerous for real-world applications such as driving scene perception where diverse objects and unexpected behaviours should be expected. In order to progress towards trusted autonomous platforms object recognition models need Open Set Recognition (OSR) methods capable of identifying unknown classes while maintaining good performance on known classes. Existing OSR methods are mostly designed for image data and utilize generative models which are hard to train. In this paper, we propose S2CA, a Supervised Contrastive Class Anchor learning method which leverages contrastive learning principles to effectively reject unknown classes by increasing intra-class compactness and inter-class sparsity of known classes in feature space. We train a feature encoder through contrastive learning while ensuring that features of known classes form compact clusters, and then transfer the trained encoder to the OSR task. During inference, the model rejects unknown classes based on class-agnostic information in feature space and class-related information in logit space. The proposed OSR method is simple yet powerful. It is not only suitable for image-based object recognition models, but can also be used for a variety of lidar-based object recognition models. We demonstrate superior performance of S2CA when compared with state of the art methods on two widely used driving scene recognition datasets, i.e., KITTI and nuScenes.