MDE-VIO: Enhancing Visual-Inertial Odometry Using Learned Depth Priors

Arda Alniak; Sinan Kalkan; Mustafa Mert Ankarali; Afsar Saranli; Abdullah Aydin Alatan

cs.CVFeb 11, 2026

MDE-VIO: Enhancing Visual-Inertial Odometry Using Learned Depth Priors

Authors: Arda Alniak, Sinan Kalkan, Mustafa Mert Ankarali, Afsar Saranli, Abdullah Aydin Alatan

Organizations: Dept. of Electrical & Electronics Engineering · Center for Image Analysis (OGAM), METU, Ankara, Türkiye · Dept. of Computer Engineering

Abstract

Traditional monocular Visual-Inertial Odometry (VIO) systems struggle in low-texture environments where sparse visual features are insufficient for accurate pose estimation. To address this, dense Monocular Depth Estimation (MDE) has been widely explored as a complementary information source. While recent Vision Transformer (ViT) based complex foundational models offer dense, geometrically consistent depth, their computational demands typically preclude them from real-time edge deployment. Our work bridges this gap by integrating learned depth priors directly into the VINS-Mono optimization backend. We propose a novel framework that enforces affine-invariant depth consistency and pairwise ordinal constraints, explicitly filtering unstable artifacts via variance-based gating. This approach strictly adheres to the computational limits of edge devices while robustly recovering metric scale. Extensive experiments on the TartanGround and M3ED datasets demonstrate that our method prevents divergence in challenging scenarios and delivers significant accuracy gains, reducing Absolute Trajectory Error (ATE) by up to 28.3%. Code will be made available.

Figures & tables

Figure 1: High-level overview of the proposed MDE-VIO framework. The system augments the standard VINS-Mono by processing monocular images through a Monocular Depth Estimator. These learned depth priors are integrated into both the Front-End and Back-End of the estimator to achieve robust, metric-scaled trajectory estimation on robotic platforms.

Work	Metric	Edge Deployable	Other Limitation
[ 27 ]	✓*	✗	High compute (offline training)
[ 10 ]	✓*	✗	High compute (online training)
[ 11 ]	✓*	✗	High compute (dual complex MDEs)
[ 32 ]	✓*	✗	High compute (heavy architecture)
[ 8 ]	✓*	✓	Limited generalization
[ 21 ]	✓*	✗	High compute (heavy architecture)

Table 1: A Summary of MDE-aided optimization-based VIO systems: (*) Requires prior knowledge of the test domain to yield accurate metric depth; resulting in limited generalization.

Figure 2: MDE-VIO: The proposed VIO enhancement framework. ( Left ) The DIFT module replaces the Blue channel of the RGB input with normalized MDE predictions to enhance KLT tracking in low-texture regions. ( Right ) The proposed approach aligns learned depth priors $d_{MDE}$ for the set of tracked features $\mathcal{F}_{k}$ using affine parameters $(s,t)$ , filters unstable estimates via a variance-based gate $w_{gate}(\sigma_{i}^{2})$ , and integrates both unary depth factors $r_{\mathcal{D}}$ and pairwise OrC $r(d_{i},d_{j})$ into the VIO Optimizer.

MDE	Integration	P2001	P2002	P2003	P2004	P2005	P2006	P2007	P2008	Avg	Improvement
-	VINS-Mono (Baseline)	0.80	0.57	1.38	0.35	0.38	0.88	0.24	0.81	0.676	-
DAAC	MDI	0.68	0.53	1.24	0.35	0.37	0.75	0.26	0.83	0.626	+7.4%
	Depth Residuals + MDI	0.43	0.53	0.99	0.31	0.24	0.56	0.26	0.82	0.518	+23.4%
	OrC + MDI	0.57	0.49	1.22	0.33	0.36	0.75	0.26	0.77	0.594	+12.2%
	Depth Residuals + OrC + MDI	0.53	0.47	1.07	0.32	0.26	0.50	0.26	0.87	0.535	+20.9%
	Inverse DIFT	0.83	1.06	1.04	0.43	0.53	0.82	0.55	0.70	0.751	-10.2%

Table 2: Quantitative results TartanGround Dataset (Downtown): ATE RMSE in meters, median of 10 runs. Weights: Depth=300, Ordinal=10. Bold / underline : best/second best.

MDE	Method	Forest		Indoor			Outdoor		Summary
MDE	Method	Hard	Road	Stairs	Build	Stairwell	Penno	Skate	Avg	Improvement
-	VINS-Mono (Baseline)	0.57	0.73	X	0.31	0.53	0.60	0.15	0.482	-
DAAC	MDI	0.55	0.70	X	0.30	0.51	0.57	0.15	0.463	+3.9%
	Depth Residuals + MDI	0.54	0.65	X	0.28	0.51	0.53	0.15	0.443	+8.1%
	OrC + MDI	0.54	0.67	X	0.29	0.48	0.51	0.12	0.435	+9.8%
	Depth Residuals + OrC + MDI	0.54	0.65	X	0.28	0.55	0.50	0.15	0.445	+7.7%

Table 3: Quantitative results on M3ED: ATE RMSE in meters, median of 10 runs. Columns use aliases (e.g., ‘Penno’ for ‘penno_short_loop’). Weights: Depth=100, Ordinal=10. Bold / underline : best/second best. X denotes divergence. Note: VDA-based methods successfully track ‘Stairs’ where others diverge; summary improvement metrics exclude this sequence. VDA-based DIFT approach is not real-time feasible in this dataset (25 FPS), thus we do not present its results.

Explore similar work

AsyncMDE: Real-Time Monocular Depth Estimation via Asynchronous Spatial Memory
Mar 11, 2026Lianjie Ma, Yuquan Li, Bingzheng Jiang +3Monocular Depth Estimation Robotic Perception
School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China. · Yichang Testing Technology Research Institute, Yichang, China. · School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, China.
VIDAR: Visual-Inertial Dense Alignment and Reconstruction via a Geometric Foundation Model
Jul 19, 2026Diyari Mohammed Salih, Lingxiang Hu, Naima AitOufroukh-Mammar +1Inertial Navigation Systems Geometric Metrics
IBISC Laboratory, Université d’Évry Paris-Saclay, Université Paris-Saclay, Évry-Courcouronnes, France.
DB-VIO: Dual-Branch Visual Inertial Odometry with Enhanced Visual-Inertial Representation
Jul 24, 2026Ziyu Wan, Lin ZhaoVisual Inertial Odometry Visual-Inertial Systems
Department of Electrical and Computer Engineering, National University of Singapore, Singapore.

Mar 11, 2026, cs.ROJ/K move · Enter open · S save

AsyncMDE: Real-Time Monocular Depth Estimation via Asynchronous Spatial Memory

Lianjie Ma, Yuquan Li, Bingzheng Jiang, Ziming Zhong+2

School of Artificial Intelligence and Automation, Huazhong University of Science and Technology, Wuhan, China. · Yichang Testing Technology Research Institute, Yichang, China. · School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, China.

Monocular Depth Estimation Robotic Perception 3D Perception Fast Inference Real-Time Systems Spatial Memory

Open paper AskPDF arXiv

Foundation-model-based monocular depth estimation offers a viable alternative to active sensors for robot perception, yet its computational cost often prohibits deployment on edge platforms. Existing methods perform independent per-frame inference, wasting the substantial computational redundancy between adjacent viewpoints in continuous robot operation. This paper presents AsyncMDE, an asynchronous depth perception system consisting of a frozen foundation model and a lightweight fast path that amortizes the foundation model's computational cost over time. The foundation model periodically produces high-quality spatial features in the background, while the lightweight fast path runs asynchronously in the foreground, fusing cached memory with current observations through complementary fusion, outputting depth estimates, and autoregressively updating memory. This enables cross-frame feature reuse with bounded accuracy degradation. With 3.83M trainable fast-path parameters and a 97.5M frozen slow path, AsyncMDE's fast path operates at 237 FPS on an RTX 4090, recovering 77% of the accuracy gap to the foundation model. Across indoor static, dynamic, and synthetic extreme-motion benchmarks, AsyncMDE degrades predictably and reaches 161 FPS fast-path inference on a TensorRT-optimized Jetson AGX Orin, supporting real-time edge deployment.

Abstract

Explore similar work

AsyncMDE: Real-Time Monocular Depth Estimation via Asynchronous Spatial Memory

VIDAR: Visual-Inertial Dense Alignment and Reconstruction via a Geometric Foundation Model

DB-VIO: Dual-Branch Visual Inertial Odometry with Enhanced Visual-Inertial Representation