Improving Object Detection of Intelligent Vehicles Through Self-Training With Accurate Labeling and Class Balancing

Self-training is a novel learning paradigm that generates pseudolabels for unlabeled data, enabling deep learning models to be trained without the need for humanlabeled data. This article proposes self-training through accurate labeling and class balancing (SALB) method that enhances the pre-trained models through periodic multi-round self-training with pseudo-labeled data. In this context, we focus on generating high quality labels by predicting the maximum possible detection labels using different augmented views of the same image. We consolidate all predictions using a modified version of Weighted Box Fusion (WBF) and validate final pseudo-labels through adaptive confidence thresholding. Finally, we recover missing pseudo-labels through our bidirectional tracking technique. Due to the class imbalance in most available public training datasets, pre-trained models occasionally yield incorrect detections for minority object instances, resulting in a bias towards predicting the objects representing majority classes. To tackle this issue, we use copy-paste augmentation technique that enables the copy of minority instances from labeled or high confidence pseudolabeled data and paste them into pseudo-labeled data to ensure class balance. Our experiments prove that our self-training framework outperforms reference methods on Waymo dataset by achieving 8.7% mAP improvement of the initial pre-trained model with only 10% labeled data used during the model self-training.