Improving Visual Object Detection Using Synthetic Self-Training

Abstract

The current era of supervised learning requires a large corpus of application-specific training data with ground-truth annotations. The creation of large annotated datasets however is a costly endeavor. Moreover, the availability of a large annotated set of training data cannot be guaranteed in certain domains. Self-training attempts to overcome these problems by using a set of labeled data and a potentially infinite pool of unlabeled data to train a model in a semi-supervised manner. Self-training however only works if the annotated data is sufficient for training a strong teacher model, which depending on the application domain, is not always available. In this work, we propose and formulate a simple extension to the self-training paradigm and refer to it as Synthetic Self-Training (SST). SST is able to overcome the aforementioned problem by incorporating synthetically generated images into the training process, therefore improving model performance. Specifically, we address the problem of object detection in a logistics environment and are able to improve the state-of-the-art detection performance on the LOCO dataset by 12% mAP0.5.