Uncertainty-Aware 3D Plant Reconstruction from Sparse Video Frames Using Neural Radiance Fields

Abstract

Abstract Automated three-dimensional reconstruction of plant architecture under- pins high-throughput crop phenotyping, yet its deployment in practical field settings is constrained by two fundamental limitations of Neural Radiance Fields (NeRF): degraded geometry under sparse, unstructured image capture, and a complete absence of calibrated uncertainty estimates that would allow practitioners to distinguish reliable geometry from reconstruction artefacts. We present UA-PlantNeRF, a unified framework that resolves both limita- tions through three tightly coupled contributions. First, building on VISAR, our prior intelligent video frame-selection strategy (weights α1 = 0.4, α2 = 0.3, α3 = 0.3), the pipeline identifies maximally informative, non-redundant view- points from raw footage with as few as 15 frames. Second, a dual-head NeRF architecture augmented with Monte Carlo Dropout produces jointly decom- posed aleatoric and epistemic uncertainty alongside each reconstructed voxel, trained under a heteroscedastic negative log-likelihood objective. Third, split conformal prediction—with calibration performed on held-out plants to preserve exchangeability— yields provable, distribution-free per-ray cov- erage guarantees at any user-specified confidence level. Evaluated across three benchmarks (Pheno4D, ROSE-X, and our novel UA-Field greenhouse dataset) at sparsity levels N ∈ {15, 30, 50}, UA-PlantNeRF achieves PSNR 28.7±0.6 dB and SSIM 0.89±0.01 at N = 30, outperforming all sparse-view baselines (p

Plant phenotyping relevance

植物の3D構造・アーキテクチャを推定する画像ベースのNeRF手法を開発し、複数ベンチマークで評価しているため、植物表現型計測法が中心です。

Code and data availability

The paper describes two paper-specific assets: the UA-Field greenhouse plant reconstruction dataset (raw scans and metadata, Zenodo) and the authors' UA-PlantNeRF code repository. However, both are explicitly embargoed/unavailable until acceptance, so no public, actionable asset exists yet; authors must be contacted.

No evidence-backed public reproduction asset is currently recorded.