UCLA

Neural Radiance Fields (NeRF) have recently emerged as a promising research area due to their remarkable photorealistic rendering capabilities and diverse applications. By training NeRF network with various viewpoints and their corresponding images, we can generate new views of a scene. This approach, however, is highly dependent on the quality and composition of the training images. Effective learning in NeRF can be achieved by selecting images with high coverage rather than those having many overlaps. To date, there has been a lack of research focused on the selection of images for NeRF training. To address this gap, this paper proposes an innovative image selector designed to select optimal image viewpoints to enhance NeRF training. We created a custom dataset by creating a virtual environment in Unity, allowing for the flexible positioning of cameras to capture images. We introduce two methods to predict optimal image viewpoints: the first employs reinforcement learning, while the second utilizes a Self-supervised NeRF Image Selector (SNIS) based on a simplified deep neural network architecture. SNIS is trained using our novel pseudo-labels, which function analogously to the reward mechanism in reinforcement learning frameworks. Experimental results demonstrate that our selector significantly outperforms the random selection of images, providing more stable and superior performance in NeRF training.