| Summary: | 碩士 === 國立臺北科技大學 === 電子工程系 === 108 === Existing learning-based object detection methods often utilize huge amont of clean images and relatively few of low-quality images for training the neural networks. However, despite the good results they achieved on those clean data, it often struggled when it comes to those low-quality one because of its non-specified training policy. In this work, we present a conceptually simple and light-weighted framework for object detection in multiple domain data by using the feature domain transformation via Generative Adversarial Networks (GAN). As demonstrated in the experiment section, the proposed training policy can effectively learns the feature domain adaptation via an unsupervised manner, achieving state-of-the-art performance in the proposed multi domain dataset without any additional com- putational cost as well as network modifications. In addition, our experiment also shows the possibility of correctly applying GAN framework on feature spaces can benefits high-level computer vision tasks.
|
|---|
