Kuniaki Saito is a PhD student at Boston University and is advised by Professor Kate Saenko. His paper "Strong-Weak Distribution Alignment for Adaptive Object Detection”, proposes an unsupervised adaptation for both label-rich and label-poor domains and reduce annotation costs associated with detection.
This episode is a live recording of Saito presenting his paper during the CVPR poster session. He discussed his project in detail and how his proposed method is able to perform better than conventional domain adaptions.
My name is Kuniaki Saito. I'm from Boston University. I’d like to present our work, “Strong-Weak Distribution Alignment for Adaptive Object Detection”. My advisor is Kate Saenko. This is a joint work with Yoshitaka Ushiku and Tatsuya Harada.
The task we proposed in this work is domain adaptive object detection. The task is that we have source domain images. In this domain, we have images with annotated examples. In the target domain, we only have the images. The goal of this task is to develop an object detector that works well on the target domain.
The difficulty of the task is the differences of the domains. There are many parts of domain difference, such as the color of the images and the layer of the images can be different. Conventionally the domain adaptation is solved by adversarial running between feature extractor and domain classifier. In such method, the feature extractor is trying to fully align feature distribution between different domains. But what we found in this work is actually such full alignment can degrade the performance in domain adaptation. This is because the images from different domains have totally different layers. For example, the source domain can contain images focusing on only one object, whereas the target domain images can contain much more objects. So actually aligning feature distributions, in such cases actually degrades the performance. This is our motivation to start this work.
The way we solve this problem is we proposed weak alignment approach for domain adaptation. If we use existing domain classifier in this task, it will fully align feature distributions. But we designed a new loss to encourage weak alignment. We could align features, for example, that’s similar to the other domain. We focused on the examples similar to the other domain and we forced the feature extractor to align such features, and we achieved weak alignment. We could see great improvement through many datasets, such as adaptation between realistic images and artistic images. Also we could see the improvement in autonomous driving car dataset. The domain safety such as real image scenes to some different weather conditions, we could also see the improvement for synthetic to real adaptation.
