Research map / 3D vision / NeRF and radiance fields
Neural Field Classifiers via Target Encoding and Classification Loss
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Reformulates neural field regressors (NeRF, DVGO, D-NeRF, NeuS) as classifiers by binary-encoding 8-bit color targets into 24 logits and training with a BCE classification loss plus MSE.
Problem
Neural field methods (NeRF-style novel view synthesis and neural surface reconstruction) are trained as regression models with weak per-ray supervision, which may overfit and generalize poorly; the paper asks whether classification formulations are better.
Approach
Target Encoding maps each 8-bit color channel value into a binary-number label vector, so the final layer outputs 24 logits and the network is trained with a BCE-based classification loss (channel-wise, with an MSE term), turning any neural field regressor into a classifier with near-zero extra cost.
Key results
- On Replica with DVGO, mean PSNR rises from 17.08 (regression) to 34.11 (classification), and mean LPIPS falls from 0.507 to 0.0598.
- On Replica with NeuS, mean Chamfer-l1 drops from 30.24 to 7.689 (about 74.6% reduction) and mean F-score rises from 14.87 to 55.56.
- Training overhead is small: about +6% training time for DVGO and +4% for NeuS, with near-zero rendering overhead.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace regression with target encoding and classification loss | DVGO · Replica (8-scene mean) | PSNR: 17.08 -> 34.11 | Helped |
| Replace regression with target encoding and classification loss | DVGO, NeRF, and NeuS · T&T Advanced (4-scene mean) | PSNR: 22.41 -> 23.18; 22.16 -> 22.68; 19.97 -> 21.67 | Helped |
| Replace regression with target encoding and classification loss | D-NeRF · LEGO and Hook | PSNR: 21.64 -> 23.11; 29.25 -> 29.45 | Helped |
| Replace regression with target encoding and classification loss | DVGO · T&T Truck with sparse training images | PSNR: 14.87 -> 19.38; 18.76 -> 21.73; 21.02 -> 22.27; 21.72 -> 22.46 | Helped |
| Replace regression with target encoding and classification loss | DVGO · T&T Truck with corrupted training images | PSNR: 21.97 -> 22.33; 21.33 -> 22.08; 19.67 -> 21.45 | Helped |
| Replace regression with target encoding and classification loss | NeuS · Replica (8-scene mean) | Chamfer-ℓ1: 30.24 -> 7.689 | Helped |
| Add target encoding to NFC (ablation) | DVGO · Replica Scene 6 | PSNR: 27.99 -> 29.75 | Helped |
| Add target encoding to NFC (ablation) | NeuS · Replica Scene 7 | PSNR: 30.03 -> 31.28 | Helped |
| Add target encoding to NFC (ablation) | DVGO · Replica Scene 1 | PSNR: 32.52 -> 34.63 | Helped |
| Add target encoding to NFC (ablation) | DVGO · Replica Scene 2 | PSNR: 33.92 -> 33.82 | Hurt |
| Add target encoding to NFC (ablation) | DVGO · Replica Scene 3 | PSNR: 28.69 -> 34.04 | Helped |
| Add target encoding to NFC (ablation) | DVGO · Replica Scene 4 | PSNR: 37.35 -> 36.52 | Hurt |
Surprising findings
- The classification loss alone accounts for most of the gain; removing Target Encoding still yields large improvements on most scenes, and on some scenes (e.g., Scene 2, 4, 8) the variant without Target Encoding scores higher.
- The gains are reported to come mainly from better generalization (test PSNR) rather than better training fit.
Limitations
- Gains on Strivec and 4DGS are modest, and 4DGS uses NFC without target encoding because target encoding is expensive for Gaussian splatting.
- Evaluated mainly on 3D vision scenes with a few backbones; the authors acknowledge the framework is not yet validated beyond 3D reconstruction.
- Theoretical explanation for why classification improves generalization is left to future work.
Experimental setup
Datasets: Replica Tanks and Temples D-NeRF synthetic scenes (LEGO, Hook, Mutant) NeRF Synthetic Chair
Benchmarks: Replica Tanks and Temples D-NeRF LEGO and Hook NeRF Synthetic Chair
Metrics: PSNR SSIM LPIPS Chamfer-l1 Accuracy Completeness Precision Recall F-score Normal consistency
Baselines: DVGO NeRF D-NeRF NeuS Strivec 4DGS
Builds on: NeRF DVGO D-NeRF NeuS
Closest papers by meaning
- How to Train Neural Field Representations: A Comprehensive Study and Benchmark (CVPR 2024 · similarity 0.87)
- Learning Robust Generalizable Radiance Field with Visibility and Feature Augmented Point Representation (ICLR 2024 · similarity 0.84)
- ResFields: Residual Neural Fields for Spatiotemporal Signals (ICLR 2024 · similarity 0.84)
- Meta-Continual Learning of Neural Fields (ICLR 2025 · similarity 0.84)
- Coordinate-Aware Modulation for Neural Fields (ICLR 2024 · similarity 0.84)
- Novel Class Discovery for Representation of Real-World Heritage Data as Neural Radiance Fields (Student Abstract) (AAAI 2024 · similarity 0.83)
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.