Selective Domain Transformation Using Deep Learning for Improving Object Detection Accuracy
Seiya Okamoto, Masataka Seo · 2023
In this study, we aim to perform a high-precision domain transformation of data captured in any environment, including unfavorable ones, into data captured in a typical environment. This will improve the accuracy of various tasks, such as object detection. Our proposed method focuses on latent space distribution control when performing domain transformation. This method first forces the encoder, using instance normalization, to acquire a latent space strongly related only to the objects in the image. Furthermore, by minimizing the distribution discrepancy in the latent space between the input image and the image after domain transformation, domain information is excluded from the latent space, and subsequent domain transformation is facilitated. Furthermore, adaptive instance normalization achieves domain transformation that does not depend on the input domain. In the experiments in this paper, domain transformation was performed on road images and object detection was performed using a general model. Then, the proposed method is evaluated using this detection accuracy.