Beyond Motion: Fine-Grained Surface Change Forecasting under Limited Data

Mahule Roy, Subhas Roy

Keywords: surface change forecasting, localized texture evolution, low-motion prediction, synthetic progression augmentation, medical imaging, agricultural monitoring, industrial inspection, materials degradation, spatiotemporal modeling

TL;DR: We propose a lightweight spatiotemporal model to predict subtle, localized surface changes in images using synthetic progression augmentation, improving change localization across medical, agricultural, industrial, and materials datasets.

Abstract: We study the problem of forecasting subtle surface-level changes in image sequences, where the primary signal lies in localized texture evolution rather than object motion. Such scenarios arise in medical monitoring, agriculture, and material inspection, yet remain less explored compared to motion-centric video prediction. We formulate this task under limited-data conditions and propose a lightweight spatiotemporal model that combines spatial attention with explicit temporal difference modeling. To mitigate overfitting, we incorporate a synthetic progression augmentation strategy that simulates plausible texture evolution during training without mirroring evaluation-time simulations. Experiments on four small curated datasets—including real longitudinal medical and material sequences and domain-inspired simulated agricultural and industrial progression data—show modest but consistent improvements over adapted video prediction baselines, particularly in localizing regions of change. While performance remains constrained by dataset size and variability, our results suggest that explicitly modeling fine-grained texture evolution improves forecasting in non-motion settings. This work provides an initial empirical exploration of surface change forecasting as a complementary direction within visual precognition research.

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 1