Research map / 3D vision / NeRF and radiance fields
Uncertainty modeling for fine-tuned implicit functions
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.
Proposes Dropsembles, which fine-tunes multiple dropout-sampled thinned copies of a dropout-pretrained implicit decoder (with optional EWC regularization) to get ensemble-like uncertainty at much lower pretraining cost.
Problem
Uncertainty estimation for implicit functions (NeRFs, occupancy networks, SDFs) fine-tuned from dense synthetic priors on sparse, corrupted real data, where standard MC dropout produces unrealistic shapes and deep ensembles are computationally expensive.
Approach
Train one decoder with dropout on the synthetic source dataset, then sample several thinned networks from it and fine-tune each independently on the target data, treating them as a uniform ensemble; EWC regularization toward the source weights is added to limit drift from the prior.
Key results
- On the lumbar spine corrupted atlas, Dropsembles + EWC reaches 86.9% DSC versus 86.5% for Ensembles and 84.8% for MC dropout, at 48 pretraining GPU-hours versus 192 for Ensembles.
- On the toy classification task, Dropsembles + EWC attains 96.2% Acc-A and 93.5% Acc-B, comparable to Ensembles + EWC (95.8% and 95.5%).
- On ShapeNet with SIREN activations, Dropsembles + EWC achieves the best mDICE (84.5%) and ECE (17.73%) among the compared methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add EWC regularization to Dropsembles (ablation) | Dropsembles · Toy classification, dataset B | accuracy: 90.5 \pm 0.6 -> 93.5 \pm 3.8 | Helped |
| add EWC regularization to Dropsembles (ablation) | Dropsembles · MNIST | DSC: 64.1 \pm 4.8 -> 70.3 \pm 1.7 | Helped |
| add EWC regularization to Dropsembles (ablation) | Dropsembles · ShapeNet | mDSC: 83.5 \pm 2.1 -> 84.3 \pm 2.1 | Helped |
| add EWC regularization to Dropsembles (ablation) | Dropsembles · Corrupted atlas | DSC: 86.8 \pm 2.4 -> 86.9 \pm 2.4 | No clear effect |
| add EWC regularization to Dropsembles (ablation) | Dropsembles · Corrupted atlas | HD: 11.4 \pm 1.8 -> 10.9 \pm 2.1 | Helped |
| use Dropsembles instead of MCdropout | fine-tuned uncertainty methods · Corrupted atlas | DSC: 83.0 \pm 2.0 -> 86.3 \pm 2.3 | Helped |
| use Dropsembles instead of Ensembles | fine-tuned uncertainty methods · Corrupted atlas | DSC: 86.0 \pm 2.9 -> 86.3 \pm 2.3 | No clear effect |
| use Dropsembles instead of Ensembles | lumbar spine MRI inference · Lumbar spine MRI, Subject 1 | inference time: 21 -> 20 | Helped |
| use Dropsembles instead of Ensembles | lumbar spine MRI inference · Lumbar spine MRI, Subject 2 | inference time: 21 -> 19 | Helped |
| replace ReLU activations with SIREN periodic activations (ablation) | Dropsembles + EWC · ShapeNet, fine-tuned | mDICE | Mixed |
Surprising findings
- MC dropout yields noisy, non-smooth shapes when its samples are used for uncertainty, even though its point predictions are reasonable.
- Dropsembles' thinned networks share correlated initialization yet match full deep ensembles' accuracy and calibration in most experiments.
Limitations
- Fine-tuning cost on the target data is the same as deep ensembles.
- Only epistemic uncertainty of the decoder is modeled; the encoder is kept frozen.
- Dropsembles' ECE is sometimes worse than MC dropout or Ensembles on lumbar spine experiments.
Experimental setup
Datasets: MNIST ShapeNet airplanes TurboSquid rigged anatomical atlas public MR+CT lumbar spine dataset (20 subjects)
Benchmarks: toy binary classification MNIST digit reconstruction ShapeNet airplane SDF lumbar spine MRI segmentation corrupted atlas
Metrics: accuracy ECE Dice score (DSC) Hausdorff distance mIoU training time (GPU-hours)
Baselines: MC dropout Deep Ensembles MC dropout + EWC Ensembles + EWC Baseline (source-only MC dropout)
Builds on: Monte Carlo dropout deep ensembles elastic weight consolidation convolutional occupancy networks DeepSDF
Closest papers by meaning
- Variational Multi-scale Representation for Estimating Uncertainty in 3D Gaussian Splatting (NeurIPS 2024 · similarity 0.85)
- Uncertainty Estimation for Pretrained Medical Image Registration Models via Transformation Equivariance (NeurIPS 2026 · similarity 0.84)
- Bayes' Rays: Uncertainty Quantification for Neural Radiance Fields (CVPR 2024 · similarity 0.83)
- Uncertainty Visualization via Low-Dimensional Posterior Projections (CVPR 2024 · similarity 0.83)
- Hierarchical Uncertainty Estimation for Learning-based Registration in Neuroimaging (ICLR 2025 · similarity 0.83)
- Unsupervised Occupancy Learning from Sparse Point Cloud (CVPR 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.