Development of Semantic Scene Conversion Model for Image-based Localization at Night

Developing an autonomous vehicle navigation system invariant to illumination change is one of the biggest challenges in vision-based localization field due to the fact that the appearance of an image becomes inconsistent under different light conditions even with the same location. In particular, the night scene images have greatest change in appearance compared to the according day scenes. Moreover, the night images do not have enough information in Image-based localization. To deal with illumination change, image conversion methods have been researched. However, these methods could lose the detail of objects and add fake objects into the output images. In this thesis, we proposed the semantic objects conversion model using the change of local semantic objects by categories at night. This enables the proposed model to obtain the detail of local semantic objects in image conversion. As a result, it is expected that the proposed model has a better result in image-based localization. Our model uses local semantic objects (i.e., traffic signs and street lamps) as categories. The model is composed of two phases as (1) instance segmentation and (2) semantic objects conversion. Instance segmentation is utilized as a detector for local semantic objects. In translation phase, the detected local semantic objects are translated from the appearance of the night image to day image. In evaluation, we prove that models using a set of paired images show higher accuracy compared to the models using a set of unpaired images. Our proposed method will be compared with pix2pix and ToDayGAN. Moreover, the result quantitatively evaluates the best matching score with a query image and the converted images using ORB matching descriptor