An Uncertainty-Adaptive Consistency Training Algorithm for Semi-Supervised Object Counting Networks
Xiaochun Mai, Yixuan Yuan · Procedia Computer Science · 2023
Object counting presents a significant challenge for applications demanding high levels of automation. In recently ears, semi-supervised counting methods have been explored to alleviate the burden of data annotation. Despite considerable progress, the state-of-the-art methods treat pixels with varying levels of difficulty equally during each training iteration, resulting in an inflexible consistency training that can lead to high supervision loss at the onset of training and potentially cause network collapse. To address this limitation, we propose an uncertainty-adaptive consistency training algorithm for semi-supervised object counting networks in this paper. Our approach leverages image colorization to construct colorization uncertainty, which can measure the difficulty of object pixels and background pixels. To ensure superior uncertainty estimation, we then enhance this colorization uncertainty by using density features to guide the learning of colorization. Furthermore, an uncertainty-adaptive consistency training (UACT) algorithm is introduced for the consistency training of density maps. By adaptively adjusting the uncertainty threshold, we prioritize harder pixels for selection, gradually incorporating easier ones into the consistency training. The effectiveness o f t he UACT is evaluated through extensive experiments on two datasets. Experimental results demonstrate that our method outperforms state-of-the-art semi-supervised counting methods.