DETECTren: Vehicle Object Detection Using Self-Supervised Learning based on Light-Weight Network for Low-Power Devices

Abstract

Vehicle object detection is a fundamental task in computer vision. Most modern classifiers and trackers are built upon the object detectors. For example, self-driving cars use object detection on low-power devices to capture the information from the surrounding environment. Currently, object detection uses a huge amount of labelled data to train the detector. Moreover, these detectors are designed for high-end hardware (i.e., GPUs) and cannot be used on low-power devices. In this paper, we propose DETECTren, a novel object detector that uses self-supervised learning to leverage both the limited labelled data and the huge amount of unlabelled data. DETECTren learns to accurately detect the vehicle and its bounding box. DETECTren is divided into two tasks, (1) The pretext task and (2) The downstream task. In the pretext task, DETECTren uses an autoencoder with ResNet50 as a backbone sub-network to learn rotation-invariant features. The input image is rotated three times; a 90 degree rotation of the original, a 180 degree rotation of the original, and a 270 degree rotation of the original. These three images along with the original image are fed into the pretext sub-network to output a rotation invariant image. In the downstream task, a detector sub-network is used to detect and regress the bounding boxes coordinates. To match the output of the pretext task and the input of the downstream task, matching convolutional layers layers are used with trainable parameters. DETECTren is implemented using mixed-precision to be compatible with low-power devices. Experiments on the Kitti dataset show that DETECTren achieves high Average Precision (AP).