Fig. 1. Overview of training NeRF using light field data. By sampling light field 5D coordinates (position and viewing directions) along camera rays, feeding these positions into MLP to produce color and volume densities, and using volume rendering techniques to synthesize these values into corresponding image. Rendering losses are fitted using gradient descent to optimize scene representation

Fig. 2. Process of NeRF based light field super-resolution in angular domain. (a) Input 5*5 sub-aperture image; (b) expression of optical field using NeRF; (c) output 9*9 sub-aperture image

Fig. 3. Diagram of NeRF training network structure. MLP takes x (3D coordinate points and view directions) as input through first eight fully connected layers (FC), after which feature vectors will be connected to position encoding γ(d) of input observation directions, with output σ and RGB values

Fig. 4. New perspective for angular super-resolution generation of HCI antinous scenes using different methods. (a) Ground truth; (b) LFEPICNN method; (c) LLFF method; (d) proposed method; (e), (f), (g), (h) local zooms of antinous scene after using above methods, respectively

Fig. 5. New perspective for angular super-resolution generation of HCI boardgames scenes using different methods. (a) Ground truth; (b) LFEPICNN method; (c) LLFF method; (d) proposed method; (e), (f), (g), (h) local zooms of boardgames scene after using above methods, respectively

Fig. 6. Comparison of partial EPI before and after angular super-resolution of HCI boardgames scene. (a) Fix u=3, y=200 to draw EPI; (b) partial EPI of original data; (c) partial EPI of position corresponding to original data after angular super-resolution; (d) partial EPI after angular super-resolution

Fig. 7. New perspective for angular super-resolution generation of Stanford knights scenes. (a) Ground truth; (b) LFEPICNN method; (c) LLFF method; (d) proposed method; (e), (f), (g), (h) local zooms of knights scene after using above methods, respectively

Fig. 8. New perspective for angular super-resolution generation of Stanford bunny scenes. (a) Ground truth; (b) LFEPICNN method; (c) LLFF method; (d) proposed method; (e), (f), (g), (h) local zooms of bunny scene after using above methods, respectively

Fig. 9. Comparison of partial EPI before and after angular super-resolution of Stanford knights scene. (a) Fix u=3, y=210 to draw EPI; (b) partial EPI of original data; (c) partial EPI of position corresponding to original data after angular super-resolution; (d) partial EPI after angular super-resolution