We propose pixelNeRF, a learning framework that predicts a continuous neural scene representation conditioned on one or few input images. The existing approach for constructing neural radiance fields [Mildenhall et al. 2020] involves optimizing the representation to every scene independently, requiring many calibrated views and significant compute time. We take a step towards resolving these shortcomings by introducing an architecture that conditions a NeRF on image inputs in a fully convolutional manner. This allows the network to be trained across multiple scenes to learn a scene prior, enabling it to perform novel view synthesis in a feed-forward manner from a sparse set of views (as few as one).
Leveraging the volume rendering approach of NeRF, our model can be trained directly from images with no explicit 3D supervision. We conduct extensive experiments on ShapeNet benchmarks for single image novel view synthesis tasks with held-out objects as well as entire unseen categories. We further demonstrate the flexibility of pixelNeRF by demonstrating it on multi-object ShapeNet scenes and real scenes from the DTU dataset. In all cases, pixelNeRF outperforms current state-of-the-art baselines for novel view synthesis and single image 3D reconstruction.
Using multiview image supervision, we train a single pixelNeRF to 13 largest object categories in ShapeNet in order to perform novel-view synthesis on unseen objects. Our approach operates in view-space—as opposed to canonical—and requires no test-time optimization. Nevertheless, in terms of image metrics, we significantly outperform existing methods quantitatively, as shown in the paper.
Since our method requires neither canonical space nor object-level information such as masks, it can represent scenes with multiple objects, where a canonical space is unavailable, without modification. Our method can also seemlessly integrate multiple views at test-time to obtain better results. SRN performs extremely poorly here due to the lack of a consistent canonical space.
We show that our method can also conduct wide-baseline view synthesis on more complex real scenes from the DTU MVS dataset, producing reasonable results when given only 1-3 views at inference time. Moreover, it is feed-forward without requiring test-time optimization for each scene.
To demonstrate generalization capabilities, we apply a model trained on ShapeNet planes, cars, and chairs to unseen ShapeNet categories.
Separately, we apply a pretrained model on real car images after background removal.