Neural View Synthesis (NVS) is an exciting and complex challenge that can generate realistic 3D scenes from multi-view videos even in diverse real-world scenarios. To push the boundaries of NVS capabilities, a team of researchers from Purdue University, Adobe, Rutgers University and Google developed the DL3DV-140 benchmark as a litmus test for the effectiveness of NVS techniques. This benchmark is derived from DL3DV-10K, a large-scale multi-view scene dataset, which is strategically designed to capture variations in environmental settings, lighting conditions, reflective surfaces, and transparent materials. The experiment results revealed that Zip-NeRF, Mip-NeRF 360, and 3DGS are the top three performers in terms of Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). Additionally, DL3DV-10K was demonstrated to improve the performance of a state-of-the-art method, IBRNet, in terms of generalizability across various benchmarks.
This ground-breaking research not only benchmarks existing methods but also encourages exploration of DL3DV-10K to train generalizable NeRFs. This work opens up a world of possibilities and sets the foundations for further advancements in the field of Neural View Synthesis. The researchers’ efforts have the potential to revolutionise 3D representation learning and its applications. We are thrilled to witness this research and the potential impact it can have in the NVS field!
