Self-Supervised Learning for Object Detection in Autonomous Driving

Authors: Daniel Pototzky, Azhar Sultan, Matthias Kirschner, Lars Schmidt-Thieme

Abstract: Recently, self-supervised pretraining methods have achieved impressive results, matching ImageNet weights on a variety of downstream tasks including object detection. Despite their success, these methods have some limitations. Most of them are optimized for image classification and compute only a global feature vector describing an entire image. On top of that, they rely on large batch sizes, a huge amount of unlabeled data, and vast computing resources to work well.
To address these issues, we propose SSLAD, a Self-Supervised Learning approach especially suited for object detection in the context of Autonomous Driving. SSLAD computes local image descriptors that are invariant to augmentations and scale. In experiments, we show that our method outperforms state-of-the-art self-supervised pretraining methods, on various object detection datasets from the automotive domain. By leveraging just 20,000 unlabeled images taken from a video camera on a car, SSLAD almost matches ImageNet weights trained on 1,200,000 labeled images. If as few as $20$ unlabeled images are available, SSLAD generates weights that are far better than a random initialization, whereas competing self-supervised methods just do not work given so little data. Furthermore, SSLAD is very robust with respect to batch size. Even with a batch size of one, SSLAD generates weights that are clearly superior to a random initialization while greatly outperforming other self-supervised methods. This property of SSLAD even allows for single-GPU training with only a minor decrease in performance.