Depth Refinement with Sparse Geometry for Agricultural and Forest Scenes

Abstract

Deep learning-based depth estimation methods are widely studied and increasingly used in practice. They work well when test data resemble the training data, but their accuracy degrades in natural environments, such as forest and agricultural scenes, limiting their direct use for dense 3D maps in these settings. We address this problem with a simple, training-free local-affine refinement that corrects the output of a pre-trained model (e.g., stereo foundation models). Instead of retraining or fine-tuning the large number of network parameters on target data, our key idea is to use sparse, accurate geometry points, readily available from VO/VSLAM or SfM, to locally adjust the predicted depth. On agricultural and forest datasets, the proposed approach significantly improves the zero-shot performance of learned models, reaching near-centimeter accuracy at short range. At longer ranges, it surpasses a strong dense-mapping baseline (COLMAP). Overall, this lightweight, geometry-guided correction makes learned depth more reliable in challenging outdoor scenes while keeping the original network unchanged.