Training Deep Stereo Matching Networks on Tree Branch Imagery: A Benchmark Study for Real-Time UAV Forestry Applications

A new computer vision study provides a crucial benchmark for deploying AI-powered autonomous drones in forestry. Researchers from Victoria University of Wellington and Lincoln Agritech trained and tested ten state-of-the-art deep stereo matching networks specifically on real tree branch imagery, a critical step for enabling precise, real-time depth estimation needed for automated pruning. The team used the Canterbury Tree Branches dataset—5,313 stereo image pairs captured with a ZED Mini camera—and employed DEFOM-Stereo generated disparity maps as training targets.

The evaluation, conducted on an NVIDIA Jetson Orin Super module mounted on a drone, yielded clear performance trade-offs. BANet-3D delivered the highest overall perceptual quality with an SSIM score of 0.883 and LPIPS of 0.157. RAFT-Stereo scored best for high-level scene understanding with a ViTScore of 0.799. However, for practical deployment, speed was paramount. AnyNet emerged as the only model capable of near-real-time processing at 6.99 frames per second (FPS) when handling 1080P resolution, while BANet-2D offered the best quality-speed balance at 1.21 FPS.

This research directly addresses a major bottleneck in agricultural robotics: accurate, real-time 3D perception in complex, unstructured environments like forests. Small errors in disparity maps lead to significant depth mistakes, which can cause failed pruning cuts or collisions. By rigorously benchmarking models on real-world data and edge hardware, the study provides a practical roadmap for developers. It also offers guidance on resolution choices, comparing 720P and 1080P processing times to help engineers design efficient, effective forestry drone systems that can operate autonomously.