Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Towards Gold-Standard Depth Estimation for Tree Branches in UAV Forestry: Benchmarking Deep Stereo Matching Methods

Yida Lin, Bing Xue, Mengjie Zhang, Sam Schofield, Richard Green

TL;DR

This work addresses the critical need for robust, cross-domain depth estimation in vegetation-dense UAV forestry. By conducting a zero-shot comparison of eight deep stereo methods across four standard benchmarks and a new Tree Branches dataset, it reveals clear scene-dependent patterns: foundation models like DEFOM deliver superior cross-domain consistency, while iterative methods remain reliable but variable. DEFOM is proposed as the gold-standard baseline for generating pseudo-ground-truth in forestry, enabling quantitative benchmarking without LiDAR annotations. The Tree Branches dataset and DEFOM-based pseudo-ground-truth are intended to accelerate development of centimeter-level depth estimation for autonomous UAV pruning and related forestry applications.

Abstract

Autonomous UAV forestry operations require robust depth estimation with strong cross-domain generalization, yet existing evaluations focus on urban and indoor scenarios, leaving a critical gap for vegetation-dense environments. We present the first systematic zero-shot evaluation of eight stereo methods spanning iterative refinement, foundation model, diffusion-based, and 3D CNN paradigms. All methods use officially released pretrained weights (trained on Scene Flow) and are evaluated on four standard benchmarks (ETH3D, KITTI 2012/2015, Middlebury) plus a novel 5,313-pair Canterbury Tree Branches dataset ($1920 \times 1080$). Results reveal scene-dependent patterns: foundation models excel on structured scenes (BridgeDepth: 0.23 px on ETH3D; DEFOM: 4.65 px on Middlebury), while iterative methods show variable cross-benchmark performance (IGEV++: 0.36 px on ETH3D but 6.77 px on Middlebury; IGEV: 0.33 px on ETH3D but 4.99 px on Middlebury). Qualitative evaluation on the Tree Branches dataset establishes DEFOM as the gold-standard baseline for vegetation depth estimation, with superior cross-domain consistency (consistently ranking 1st-2nd across benchmarks, average rank 1.75). DEFOM predictions will serve as pseudo-ground-truth for future benchmarking.

Towards Gold-Standard Depth Estimation for Tree Branches in UAV Forestry: Benchmarking Deep Stereo Matching Methods

TL;DR

This work addresses the critical need for robust, cross-domain depth estimation in vegetation-dense UAV forestry. By conducting a zero-shot comparison of eight deep stereo methods across four standard benchmarks and a new Tree Branches dataset, it reveals clear scene-dependent patterns: foundation models like DEFOM deliver superior cross-domain consistency, while iterative methods remain reliable but variable. DEFOM is proposed as the gold-standard baseline for generating pseudo-ground-truth in forestry, enabling quantitative benchmarking without LiDAR annotations. The Tree Branches dataset and DEFOM-based pseudo-ground-truth are intended to accelerate development of centimeter-level depth estimation for autonomous UAV pruning and related forestry applications.

Abstract

Autonomous UAV forestry operations require robust depth estimation with strong cross-domain generalization, yet existing evaluations focus on urban and indoor scenarios, leaving a critical gap for vegetation-dense environments. We present the first systematic zero-shot evaluation of eight stereo methods spanning iterative refinement, foundation model, diffusion-based, and 3D CNN paradigms. All methods use officially released pretrained weights (trained on Scene Flow) and are evaluated on four standard benchmarks (ETH3D, KITTI 2012/2015, Middlebury) plus a novel 5,313-pair Canterbury Tree Branches dataset (). Results reveal scene-dependent patterns: foundation models excel on structured scenes (BridgeDepth: 0.23 px on ETH3D; DEFOM: 4.65 px on Middlebury), while iterative methods show variable cross-benchmark performance (IGEV++: 0.36 px on ETH3D but 6.77 px on Middlebury; IGEV: 0.33 px on ETH3D but 4.99 px on Middlebury). Qualitative evaluation on the Tree Branches dataset establishes DEFOM as the gold-standard baseline for vegetation depth estimation, with superior cross-domain consistency (consistently ranking 1st-2nd across benchmarks, average rank 1.75). DEFOM predictions will serve as pseudo-ground-truth for future benchmarking.
Paper Structure (23 sections, 3 equations, 4 figures, 2 tables)

This paper contains 23 sections, 3 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Deep Stereo Matching Architectures
  4. Foundation Models and Diffusion-Based Methods
  5. Stereo Vision for UAV Forestry
  6. Zero-Shot Cross-Domain Generalization
  7. Methodology
  8. Problem Formulation
  9. Datasets
  10. Evaluated Methods
  11. Evaluation Metrics
  12. Implementation Details
  13. Experimental Results
  14. Performance on Standard Benchmarks
  15. Performance Analysis by Method
  16. ...and 8 more sections

Figures (4)

  • Figure 1: Initial screening of 20 stereo matching methods using officially released pretrained weights on KITTI 2015 (D1-all %) and Middlebury (Average Absolute Error, pixels). Foundation models (DEFOM: 0.79% D1, BridgeDepth: 1.01% D1) dominate KITTI 2015, while iterative methods (IGEV++: 0.97 px AAE, DLNR: 1.06 px) excel on Middlebury. Based on this screening, eight methods (highlighted) are selected for comprehensive cross-domain generalization evaluation, balancing performance and architectural diversity. Classical 3D CNN methods (ACVNet, PSMNet) are included as baselines despite higher errors.
  • Figure 2: Qualitative comparison on Scene 3305. DEFOM produces smoother disparity maps with better sky-region consistency. IGEV++ preserves finer branch details but exhibits noise in homogeneous areas.
  • Figure 3: Qualitative comparison on Scene 4939. DEFOM generates smooth depth transitions at occlusion boundaries. IGEV++ produces sharper edges but exhibits artifacts near thin structures.
  • Figure 4: Qualitative comparison on Scene 5128 under variable lighting conditions. DEFOM maintains stable depth predictions. IGEV++ shows more texture detail but less consistent depth.