Paper summary: FrugalNeRF: Fast Convergence for Extreme Few-shot Novel View Synthesis without Learned Priors

Problem

NeRFs trained from only a handful of views overfit and take a long time to train. Existing fixes either regularise frequencies with complex schedules (FreeNeRF) or rely on pretrained priors (SparseNeRF), which can bring bias.

Method

FrugalNeRF:

• represents the scene with weight-sharing voxels at multiple scales, which captures details efficiently
• uses cross-scale geometric adaptation: at each scale it renders depth and checks reprojection error against other views, then picks the best as pseudo ground-truth depth to guide training

This makes full use of the few training images without external learned priors. Priors can still be added for extra quality without slowing convergence.

Key results

From the abstract: on LLFF, DTU and RealEstate-10K, FrugalNeRF outperforms other few-shot NeRF methods while significantly reducing training time.

Limitations

There’s no dedicated limitations section in the extracted text. Results cover these three standard few-shot benchmarks.