View Correspondence Network for Implicit Light Field Representation

• DOI: 10.48550/arxiv.2305.06233
• Publication Date: 2023-01-01
• Authors: Süleyman Aslan; Brandon Y. Feng; Amitabh Varshney
• Keywords: Light Field; Leverage (statistics); Representation; Correspondence problem

We present a novel technique for implicit neural representation of light fields at continuously defined viewpoints with high quality and fidelity. Our maps 4D coordinates defining two-plane parameterization the to corresponding color values. leverage periodic activations achieve expressivity accurate reconstruction complex data manifolds while keeping low storage inference time requirements. However, na\"ively trained non-3D structured networks do not adequately satisfy multi-view consistency; instead, they perform alpha blending nearby viewpoints. In contrast, our View Correspondence Network, or VICON, leverages stereo matching, optimization by automatic differentiation respect input space, pixel correspondence provide faithful views that are unseen during training. Experimental results show VICON superior state-of-the-art field representations both qualitatively quantitatively. Moreover, captures larger view (FoV), surpassing extent observable scene cameras ground truth renderings.