Cross-domain object detection using unsupervised image translation and neural style transfer
Nenhuma Miniatura disponível
Data
2022-02-24
Autores
Arruda, Vinicius Ferraço
Título da Revista
ISSN da Revista
Título de Volume
Editor
Universidade Federal do Espírito Santo
Resumo
Unsupervised domain adaptation for object detection addresses the adaptation of detectors trained in a source domain to work accurately in an unseen target domain. In real-world applications, object detectors are desired to work accurately regardless of the application domain (e.g., weather condition). These models have the intrinsic property of being biased towards the training data and are known to not generalize well to unseen data. The greatest availability of datasets can be seen in the most prevalent domains (e.g., sunny day), but for certain applications it may be necessary to train a model to deploy in a less prevalent one (e.g., foggy day). In addition, the acquisition of a new dataset involves the laborious process of data annotation, but collecting large amounts of data without annotation might be feasible. Recently, methods for unsupervised domain adaptation approaching the alignment of the intermediate features proven to be promising, achieving state-of-the-art results. However, these methods are laborious to implement and hard to interpret. Although promising, there is still room for improvements to close the performance gap toward the upper-bound (when training with the target data). In this work, we propose a method to generate an artificial dataset in the target domain to train an object detector. We employed an unsupervised image translator (CycleGAN) and a neural style transfer method (AdaIN-based) using only annotated data from the source domain and non-annotated data from the target domain. Our key contributions are the proposal of a less complex yet more effective method that also has an improved interpretability. Results on real-world scenarios for autonomous driving show significant improvements, outperforming state-of-the-art methods in most cases, further closing the gap toward the upper-bound.
Descrição
Palavras-chave
Adaptação de domínio não supervisionado , Redes generativas adversariais , Transferência de estilo neural