cs.ROApr 8, 2026

CADENCE: Context-Adaptive Depth Estimation for Navigation and Computational Efficiency

Authors: Timothy K Johnsen, Marco Levorato

Organizations: Computational Science joint program with San Diego State University and The University of California, Irvine in CA, USA · Computer Science department at the University of California, Irvine in CA, USA

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Autonomous vehicles deployed in remote environments typically rely on embedded processors, compact batteries, and lightweight sensors. These hardware limitations conflict with the need to derive robust representations of the environment, which often requires executing computationally intensive deep neural networks for perception. To address this challenge, we present CADENCE, an adaptive system that dynamically scales the computational complexity of a slimmable monocular depth estimation network in response to navigation needs and environmental context. By closing the loop between perception fidelity and actuation requirements, CADENCE ensures high-precision computing is only used when mission-critical. We conduct evaluations on our released open-source testbed that integrates Microsoft AirSim with an NVIDIA Jetson Orin Nano. As compared to a state-of-the-art static approach, CADENCE decreases sensor acquisitions, power consumption, and inference latency by 9.67%, 16.1%, and 74.8%, respectively. The results demonstrate an overall reduction in energy expenditure by 75.0%, along with an increase in navigation accuracy by 7.43%.

Figures & tables

Fig. 1: System environment that contains an autonomous drone equipped with an embedded computer, GPS, distance sensor, and monocular camera.
Fig. 2: Illustrated is the flow of data from raw sensor acquisition to intelligent decision-making through the full autonomy stack, CADENCE.
α\alphapower [mW]latency [ms]energy [mJ]
25617740117.12078.1
1281835130.0550.0
641780811.0196.4
32129616.583.9
1673757.051.6
861236.741.1
TABLE I:
Fig. 3: Example RGB image, ground truth depth map, and predicted depth maps for both static and slimmable networks with varying network sizes.
Fig. 4: Test R2R^{2} -scores of various trained MDE network configurations.
Fig. 5: Learning curve from training the navigation-and-adaptation policy.
Fig. 6: Final evaluations on the static holdout test set of ground truth A* paths, for each trained policy. The x-axis shows the energy expended while computing the MDE network over all paths. The hollow blue markers correspond to using a SoTA static MDE network with size α\alpha , and a navigation DQN policy that predicts aa . The solid orange markers correspond to using a slimmable MDE network with a navigation-and-adaptation DQN that predicts both aa and ρ\rho . The green dashed line corresponds to a navigation DQN that bypasses MDE and instead directly uses the AirSim ground truth depth maps.
Fig. 7: Average slimming factor when computing the slimmable MDE network at each of the given (x, y) locations. Compared is the drone moving in an arbitrary direction versus moving only to the left.
Fig. 8: Correlation between the adaptation parameter ρ\rho and two environmental factors: obstacle density, as measured by the mean value of the FIFO queue of recent depth map predictions, and distance to target.
More in this topic

Power-Efficient Inference

See all 14 papers

Explore similar work

CardsList

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

    Mar 11, 2026Lianjie Ma, Yuquan Li, Bingzheng Jiang +3Edge InferenceMonocular Depth Estimation

    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.

  2. JustDepth: Real-Time Radar-Camera Depth Estimation with Single-Scan LiDAR Supervision

    Jul 24, 2026Wooyung Yun, Dongwook Kim, Soomok LeeCamera-Radar FusionMetric Depth Estimation

    Department of Artificial Intelligence, Ajou University, Suwon 16499, Republic of Korea · Faculty of Department of Computer Science, Kennesaw State University, Marietta GA, 30060, United States

  3. An active-learning framework for real-time depth perception from monocular vision streams

    Aug 5, 2026Xiaorong Zeng, Weiqiang Chen, Peng Shi +4Edge InferenceMonocular Depth Estimation

    School of Mechanical and Automotive Engineering, Xiamen University of Technology, Xiamen, Fujian 361024, China · King Long United Automotive Industry Co. Ltd., Xiamen 361023, China · School of Electrical and Electronic Engineering, The University of Adelaide, Adelaide, SA 5005, Australia +2