Conformal Prediction-Driven Frame Selection for Resource-Constrained Agricultural 3D Plant Reconstruction

Abstract

Abstract High-fidelity 3D reconstruction of plants from video is a key enabler of digital phenotyping, but the dense frame streams produced by modern cameras impose prohibitive compute, memory, and energy costs on the embedded platforms used in the field. Existing frame-selection methods reduce this load using geometric or photometric heuristics, yet they provide no statistical guarantee on the reconstruction quality that the retained subset will deliver. We present a conformal predictiondriven frame selection strategy that augments heuristic selection with a calibrated, distribution-free decision rule: a frame is processed only when the reconstruction model’s predicted uncertainty for that view exceeds a threshold whose miscoverage rate is controlled at a user-specified level α. Because conformal prediction makes no assumption on the underlying error distribution, the resulting frame budget carries a finite-sample guarantee on the probability that reconstruction error stays below the target tolerance. On a multi-view plant dataset, the proposed method attains reconstruction fidelity comparable to processing the full stream while using 41% fewer frames, and reduces the frame budget by 23% relative to a strong uncertainty-agnostic baseline at matched quality. The approach is lightweight enough for on-device deployment and turns frame selection from a tuned heuristic into a procedure with explicit, controllable risk—a property we argue is essential for trustworthy edge agriculture.

Code and data availability

The supplied article blocks describe a conformal prediction-driven frame selection method evaluated on a multi-view plant capture dataset, but contain no data availability statement, no public dataset/repository URL, no code deposit, and no supplement reference. The dataset and pipeline are described only abstractly,so

No evidence-backed public reproduction asset is currently recorded.