Neural Plenoptic Sampling: Capture Light-field from Imaginary Eyes

The Plenoptic function describes the light rays observed from any given position in every viewing direction. It is often parameterized as a 5-D function $L(x, y, z, \theta, \phi)$ for a static scene. Capturing all the plenoptic functions in the space of interest is paramount for Image-Based Rendering (IBR) and Novel View Synthesis (NVS). It encodes a complete light-field (\ie, lumigraph) therefore allows one to freely roam in the space and view the scene from any location in any direction. However, achieving this goal by conventional light-field capture technique is expensive, requiring densely sampling the ray space using arrays of cameras or lenses. This paper proposes a much simpler solution to address this challenge by using only a small number of sparsely configured camera views as input. Specifically, we adopt a simple Multi-Layer Perceptron (MLP) network as a universal function approximator to learn the plenoptic function at every position in the space of interest. By placing virtual viewpoints (dubbed `imaginary eyes') at thousands of randomly sampled locations and leveraging multi-view geometric relationship, we train the MLP to regress the plenoptic function for the space. Our network is trained on a per-scene basis, and the training time is relatively short (in the order of tens of minutes). When the model is converged, we can freely render novel images. Extensive experiments demonstrate that our method well approximates the complete plenoptic function and generates high-quality results.