Research map / 3D vision / NeRF and radiance fields
Enhancing NeRF akin to Enhancing LLMs: Generalizable NeRF Transformer with Mixture-of-View-Experts
Wenyan Cong, Hanxue Liang, Peihao Wang, Zhiwen Fan, Tianlong Chen, Mukund Varma, Yi Wang, Zhangyang Wang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| plug a naive MoE layer into GNT view transformer without PE or smoothness loss (ablation) | GNT · LLFF (zero-shot) | PSNR: 25.86 -> 25.46 | Hurt |
| plug a naive MoE layer into GNT view transformer without PE or smoothness loss (ablation) | GNT · NeRF Synthetic (zero-shot) | PSNR: 27.29 -> 27.15 | Hurt |
| add shared permanent expert to MoE (ablation) | GNT + MoE · LLFF (zero-shot) | PSNR: 25.46 -> 25.88 | Improved |
| add geometry-aware smoothness regularizer to MoE (ablation) | GNT + MoE · LLFF (zero-shot) | PSNR: 25.46 -> 25.93 | Improved |
| add shared permanent expert to MoE (ablation) | GNT + MoE · NeRF Synthetic (zero-shot) | PSNR: 27.15 -> 27.32 | Improved |
| add both permanent expert and smoothness regularizer (ablation) | GNT + MoE · LLFF (zero-shot) | PSNR: 25.46 -> 26.02 | Improved |
| use MoE with 2.5x total parameters instead of widening dense GNT by 2.5x (ablation) | GNT · Tanks-and-Temples Truck (zero-shot) | PSNR: 18.26 -> 19.71 | Improved |
| replace permanent expert with an extra selectable expert at equal params and FLOPs (ablation) | GNT-MOVE · Tanks-and-Temples Truck (zero-shot) | PSNR: 19.71 -> 18.11 | Hurt |
| replace GNT with GNT-MOVE on challenging unbounded scenes | GNT · Tanks-and-Temples (4 scenes, zero-shot) | PSNR: up to 3 dB PSNR | Improved |
| replace GNT with GNT-MOVE in few-shot finetuning | GNT · LLFF few-shot | PSNR: +0.13 dB (3-shot), +0.17 dB (6-shot), +0.47 dB (10-shot) | Improved |
| replace GNT with GNT-MOVE on LLFF and NeRF Synthetic SSIM versus GPNR | GNT-MOVE · LLFF and NeRF Synthetic (zero-shot) | SSIM: LLFF 0.880 (GPNR) vs 0.869; NeRF Synthetic 0.944 (GPNR) vs 0.940 | Mixed |
Closest papers by meaning
- Is Attention All That NeRF Needs? (ICLR 2023 · similarity 0.91)
- NeRFDeformer: NeRF Transformation from a Single View via 3D Scene Flows (CVPR 2024 · similarity 0.89)
- InsertNeRF: Instilling Generalizability into NeRF with HyperNet Modules (ICLR 2024 · similarity 0.89)
- NeRF Analogies: Example-Based Visual Attribute Transfer for NeRFs (CVPR 2024 · similarity 0.88)
- NeRF On-the-go: Exploiting Uncertainty for Distractor-free NeRFs in the Wild (CVPR 2024 · similarity 0.88)
- Language-driven Object Fusion into Neural Radiance Fields with Pose-Conditioned Dataset Updates (CVPR 2024 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
