Density-Based Vehicle Counting with Unsupervised Scale Selection

A significant hurdle within any counting task is the variance in scale of the objects to be counted. While size changes of some extent can be induced by perspective distortion, more severe scale differences can easily occur, e.g. in case of images taken by a drone from different elevations above the ground. The aim of our work is to overcome this issue by leveraging only lightweight dot annotations and a minimum level of training supervision. We propose a modification to the Stacked Hourglass network which enables the model to process multiple input scales and to automatically select the most suitable candidate using a quality score. We alter the training procedure to enable learning of the quality scores while avoiding their direct supervision, and thus without requiring any additional annotation effort. We evaluate our method on three standard datasets: PUCPR+, TRANCOS and CARPK. The obtained results are on par with current state-of-the-art methods while being more robust towards significant variations in input scale.