Neural Light Field 뷰 합성 성능 예측에 효과적인 이미지 복잡도 평가지표에 관한 연구

Neural Light Field (NeLF) is an approach that learns the relationship between implicit representations and the color information of light rays, using this knowledge to reconstruct real spaces or objects. NeLF has been actively researched recently due to its impressive performance. However, the training performance of NeLF varies depending on the complexity of the target space or object. In general, the training performance of NeLF for complex spaces is low, and for simple spaces it is high. Regardless of this trend, it is difficult to predict the level of PSNR before training. This paper studies methods to predict the training performance of NeLF based on the images used for training, and in particular studies image complexity metrics that are effective in predicting NeLF training performance. This paper selects six evaluation metrics used to assess image complexity and analyzes the relationship between the actual PSNR results and the corresponding complexity. The experiment is based on a total of 30 samples, including standard data samples commonly used in NeLF research as well as samples directly captured for this study, to explore the most relevant image complexity metrics for NeLF performance. The experiment also predicts PSNR for five test samples and analyzes the prediction errors.