EpiTransfer: Sparse, Training-Free Long-Range Depth Estimation from Temporal Monocular Aerial Frames

Diksha Aggarwal; Rutvik Dagadkhair; Sanjana Srivastava; Bradley Denby; Kevin Kochersberger

cs.ROSep 27, 2026

EpiTransfer: Sparse, Training-Free Long-Range Depth Estimation from Temporal Monocular Aerial Frames

Authors: Diksha Aggarwal, Rutvik Dagadkhair, Sanjana Srivastava, Bradley Denby, Kevin Kochersberger

Organizations: Department of Mechanical Engineering, Virginia Tech, Blacksburg, Virginia, USA. · Department of Aerospace and Ocean Engineering, Virginia Tech, Blacksburg, Virginia.

Abstract

Reliable 3D spatial understanding is essential for autonomous navigation, obstacle avoidance, and scene reconstruction. While state-of-the-art learned depth estimation techniques achieve high accuracy in-distribution, they often generalize poorly to novel viewpoints and altitudes. This paper presents a geometrically derived, training-free depth estimation method using epipolar transfer with only two monocular images and camera pose estimates. By leveraging camera motion to synthesize a virtual stereo pair with a freely chosen baseline, our approach transforms temporal correspondence into a stereo triangulation task while mitigating geometric degeneracies inherent to direct two-view triangulation. Validated across outdoor drone flights (to a maximum range of approximately 90,m) and indoor OptiTrack environments against LiDAR ground truth, the method achieves an indoor AbsRel of 0.092 and $δ< 1.25$ of 0.940, comparable to direct triangulation (AbsRel 0.073) while retaining valid depth over a larger fraction of challenging scenes, and substantially outperforms off-the-shelf learning-based baselines such as ZoeDepth (AbsRel 0.225) and Depth Anything V2 (AbsRel 0.570), which are not trained or fine-tuned for this domain, with no training data required.

Figures & tables

Fig. 1 : We use temporal monocular images and relative pose of cameras, perform sparse feature matching using XFeat, synthesize a virtual stereo pair via epipolar transfer, and triangulate to obtain sparse depth.

Fig. 2 : Flow chart of the proposed method, EpiTransfer

Fig. 3 : Virtual Stereo Frame synthesis using $I_{t}$ and $I_{t-1}$

Fig. 4 : Experiment setup with camera and LiDAR mounted on the tripod in the indoor OptiTrack motion capture environment

Panel of Figure 5 — Fig. 5: Qualitative comparison of depth estimation results on 5 of the 8 outdoor images used for evaluation. Black regions in EpiTransfer and Triangulation indicate pixels without a valid depth estimate, either beyond the 100 m ground-truth range, unmatched in sparse correspondence, or masked as geometrically ill-conditioned for triangulation.

Metric	EpiTransfer (Ours)	Triangulation (DLT)	ZoeDepth	Depth Anything V2
AbsRel $\downarrow$	0.138	0.124	0.719	0.744
RMSE (m) $\downarrow$	10.255	8.731	32.610	32.959
RMSE log $\downarrow$	0.199	0.174	1.279	1.381
MAE (m) $\downarrow$	6.649	5.656	29.750	30.423
SqRel $\downarrow$	1.694	1.350	21.560	22.579
$\delta<1.25$ $\uparrow$	0.755	0.817	0.000	0.000

TABLE I : Depth estimation accuracy evaluated against LiDAR ground truth on 8 images of our outdoor dataset.

Panel of Figure 7 — Fig. 6: Qualitative comparison of depth estimation results on the indoor dataset. Black regions in EpiTransfer and Triangulation indicate pixels without a valid depth estimate, either unmatched in sparse correspondence or masked as geometrically ill-conditioned for triangulation.

Metric	EpiTransfer (Ours)	Triangulation (DLT)	ZoeDepth	Depth Anything V2
AbsRel $\downarrow$	0.092	0.073	0.225	0.570
RMSE (m) $\downarrow$	0.469	0.401	0.980	2.964
RMSE log $\downarrow$	0.118	0.102	0.235	0.482
MAE (m) $\downarrow$	0.371	0.303	0.841	2.520
SqRel $\downarrow$	0.052	0.037	0.266	1.859
$\delta<1.25$ $\uparrow$	0.940	0.943	0.601	0.223

TABLE II : Depth estimation accuracy evaluated against LiDAR ground truth on 7 images of our indoor dataset.

Panel of Figure 9 — Fig. 7 : AbsRel vs. injected pose error across 7 indoor takes. Takes 2-1 and 5-4 (which include perturbations placing the epipole inside the image) show markedly larger, non-symmetric error, consistent with the trifocal-plane degeneracy of Section V-A ; the remaining 5 takes show bounded, approximately symmetric degradation around zero injected error.

Explore similar work

Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth
May 19, 2026Viktor Kocur, Sithu Aung, Gabrielle Flood +4Monocular Depth Estimation Structure-From-Motion
Faculty of Mathematics, Physics and Informatics, Comenius University in Bratislava · Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague · Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague
ZipDepth: Bringing Lightweight Zero-Shot Monocular Depth Anywhere, on Any Device
Jul 9, 2026Fabio Tosi, Luca Bartolomei, Matteo Poggi +1Monocular Depth Estimation Mobilenetv2
University of Bologna, Bologna, Italy
Depth Anything in $360^\circ$ : Towards Scale Invariance in the Wild
Dec 28, 2025Hualie Jiang, Ziyang Song, Zhiqiang Lou +2Omnidirectional Images 3D Perception
Insta360 Research

May 19, 2026, cs.CVJ/K move · Enter open · S save

Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth

Viktor Kocur, Sithu Aung, Gabrielle Flood, Yaqing Ding+3

Faculty of Mathematics, Physics and Informatics, Comenius University in Bratislava · Visual Recognition Group, Faculty of Electrical Engineering, Czech Technical University in Prague · Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague

Monocular Depth Estimation Structure-From-Motion Ground Truth Performance Metric

Open paper AskPDF arXiv

Monocular depth estimation has improved significantly in recent years, driven by increasingly powerful models and large-scale training data. Predicted depth is increasingly used as an input signal for downstream tasks such as Structure-from-Motion (SfM), visual localization, and SLAM. However, monocular depth estimators (MDEs) are still primarily evaluated in terms of depth accuracy. Standard metrics aggregate errors globally and may not reflect the usefulness of depth for downstream geometric tasks. We therefore propose Depth2Pose, a framework for evaluating MDEs in the context of downstream tasks. By combining depth predictions with feature correspondences in depth-aware geometric solvers, we use relative camera pose estimation accuracy as a task-driven proxy for depth quality. Traditional benchmarks require dense ground truth in the form of per-pixel depth, which is expensive to obtain. In contrast, our formulation requires only camera poses, which can be estimated efficiently, e.g., using Structure-from-Motion pipelines. As a result, our framework can be applied to scenes where ground-truth depth is difficult to obtain, for example due to large scene scale or heavy occlusions (e.g., vegetated environments). Leveraging this, we introduce the D2P dataset, which contains challenging scenes outside the distribution of commonly used training data. We show that methods performing well under standard depth error metrics on existing benchmarks also perform well under our pose-based metric when evaluated on the same datasets, but do not necessarily generalize to our more challenging dataset. Finally, we provide a simple and extensible evaluation framework. The dataset and code are available at kocurvik.github.io/depth2pose.

Abstract

Explore similar work

Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth

ZipDepth: Bringing Lightweight Zero-Shot Monocular Depth Anywhere, on Any Device

Depth Anything in 360∘360^\circ360∘: Towards Scale Invariance in the Wild

Depth Anything in $360^\circ$ : Towards Scale Invariance in the Wild