Unsupervised Deep Domain Adaptation for Pedestrian Detection
It addresses domain shift in pedestrian detection for applications like surveillance and autonomous driving, but it is incremental as it builds on existing deep learning methods.
This paper tackles unsupervised domain adaptation for pedestrian detection in crowded scenes by using an iterative auto-annotation algorithm and an unsupervised regularizer, boosting recall by nearly 30% while maintaining precision and achieving state-of-the-art results on benchmarks.
This paper addresses the problem of unsupervised domain adaptation on the task of pedestrian detection in crowded scenes. First, we utilize an iterative algorithm to iteratively select and auto-annotate positive pedestrian samples with high confidence as the training samples for the target domain. Meanwhile, we also reuse negative samples from the source domain to compensate for the imbalance between the amount of positive samples and negative samples. Second, based on the deep network we also design an unsupervised regularizer to mitigate influence from data noise. More specifically, we transform the last fully connected layer into two sub-layers - an element-wise multiply layer and a sum layer, and add the unsupervised regularizer to further improve the domain adaptation accuracy. In experiments for pedestrian detection, the proposed method boosts the recall value by nearly 30% while the precision stays almost the same. Furthermore, we perform our method on standard domain adaptation benchmarks on both supervised and unsupervised settings and also achieve state-of-the-art results.