SPARF: Large-Scale Learning of 3D Sparse Radiance Fields from Few Input Images

TL;DR

This work tackles few-image novel view synthesis by learning to generate sparse radiance fields (SRFs) from partial observations. It introduces SPARF, a large-scale ShapeNet-based synthetic dataset with about 1M optimized SRFs and 17 million images across multiple voxel resolutions, and SuRFNet, a 3D sparse-convolution network that completes partial SRFs into full SRFs for high-quality rendering from novel viewpoints. The approach combines a triad of losses—density, radiance color, and perceptual rendering—along with a Q-Gaussian loss-sampling strategy to train the network end-to-end on a diverse SRF distribution. Empirical results show state-of-the-art performance in unconstrained, few-view novel view synthesis on ShapeNet, with notable generalization to out-of-distribution views and real-image testing via Co3D, demonstrating the practicality of learning 3D SRFs rather than optimizing per-scene radiance fields. SPARF and SuRFNet together provide a scalable pathway for 3D radiance-field learning and faster VR/AR rendering pipelines, at the cost of substantial compute and memory requirements that future work aims to reduce.

Abstract

Recent advances in Neural Radiance Fields (NeRFs) treat the problem of novel view synthesis as Sparse Radiance Field (SRF) optimization using sparse voxels for efficient and fast rendering (plenoxels,InstantNGP). In order to leverage machine learning and adoption of SRFs as a 3D representation, we present SPARF, a large-scale ShapeNet-based synthetic dataset for novel view synthesis consisting of $\sim$ 17 million images rendered from nearly 40,000 shapes at high resolution (400 X 400 pixels). The dataset is orders of magnitude larger than existing synthetic datasets for novel view synthesis and includes more than one million 3D-optimized radiance fields with multiple voxel resolutions. Furthermore, we propose a novel pipeline (SuRFNet) that learns to generate sparse voxel radiance fields from only few views. This is done by using the densely collected SPARF dataset and 3D sparse convolutions. SuRFNet employs partial SRFs from few/one images and a specialized SRF loss to learn to generate high-quality sparse voxel radiance fields that can be rendered from novel views. Our approach achieves state-of-the-art results in the task of unconstrained novel view synthesis based on few views on ShapeNet as compared to recent baselines. The SPARF dataset is made public with the code and models on the project website https://abdullahamdi.com/sparf/ .

Figures (32)

• Figure 1: Distribution of Radiance Fields. We treat Sparse Radiance Field (SRF) as a 3D data structure and learn the conditional generation of SRFs from few input images for the task of novel view synthesis. In order to do this, we build SPARF, a large-scale dataset of SRFs.
• Figure 2: SPARF: a Large Dataset for 3D Shapes Radiance Fields and Novel Views Synthesis.
• Figure 3: SPARF vs. other Datasets. SPARF offers a large-scale high-resolution dataset compared to other posed multi-view datasets. We show the same chair here on SPARF, SRN, and NMR (please zoom in for differences). This highlights the huge quality gap between SPARF and other ShapeNet-based datasets.
• Figure 4: SRFs: The optimized Sparse Radiance Fields in SPARF. A total of one million SRFs have been collected in SPARF, including on multiple voxel resolutions: 32 (top), 128 (middle), and 512 (bottom).
• Figure 5: Extracting 3D Meshes from SRFs. Since SPARF and SuRFNet lie on the 3D voxel's space, extracting the mesh is straightforward with one pass of MarchingCubes mcubes.