Source data-free domain adaptation for a faster R-CNN

Abstract

The existing domain adaptive object detection methods often need to carry a large number of source domain samples for domain adaptation, which is not realistic due to GPU limitations, privacy and physical memory in practical applications. To solve this problem, we propose a source data-free domain adaptive object detection method. Only unlabeled target domain data is used to optimize the source domain model so that it can work better in the target domain. Our method takes Faster R-CNN as baseline. Specifically, we first construct global class prototypes which will be updated in batch iteratively. Then based on the global class prototypes, more accurate pseudo-labels are generated for training the target model. In this way, the source and target domains are also implicitly aligned. Our contributions are 1) a prototype guided domain adaptation method which uses prototypes to mine the semantic category information without accessing the source dataset; 2) a scheme of iteratively updating global class prototype which can handle the class and sample imbalances in the training procedure and 3) a more accurate pseudo-label generation method combining semantic information and image information. On multiple public domain adaptive scenarios, our method achieves the state-of-the-art results in terms of accuracy compared with the Faster R-CNN model and some domain adaptive methods with source datasets.