Sound: Unknown Object Detection for Autonomous Driving

Existing perception algorithms for autonomous vehicles are currently limited to recognizing only the objects encountered during their training. As a result, these systems often fail to identify untrained objects encountered in realworld driving scenarios, increasing the risk of accidents. In this paper, we introduce a simple yet efficient method for detecting unknown objects in autonomous driving applications. Our approach, named SOUND, effectively recognizes samples from known classes while rejecting those from previously unseen classes. This capability is critical for autonomous vehicles to navigate complex and unpredictable environments. At the core of SOUND is a novel metric learning framework, projection softmax (p-softmax), which offers a clear geometric interpretation of the evaluation metric and simplifies training compared to existing approaches. To connect the recognition of unknown samples and the classification of known samples, we further extend the feature norm to a smooth maximum projection, bridging the metric learning and energy-based model gap. In addition, we propose a fine-tuning method that enhances feature discrimination between known and unknown classes, further improving the performance. Our experimental evaluation shows that SOUND outperforms existing models and methods by a clear margin. Experimental results on the MNIST, CIFAR, and ImageNet datasets demonstrate its effectiveness and robustness. SOUND offers substantial promise for enhancing the safety and reliability of autonomous driving systems by enabling them to generalize beyond their training data and adapt to new objects and situations encountered on the road.