UtilityRoute: Task-Aware Expert Routing for Aerial RGB-T Perception

Unmanned aerial vehicles (UAVs), as dynamic sensing platforms operating under continuously changing observation conditions, benefit from complementary red-green-blue (RGB) and thermal (T) information for object detection. The detection gains of different feature interventions vary across observations, yet existing methods struggle to assess them accurately, potentially leading to ineffective interventions and redundant computation. We present UtilityRoute, an RGB-T front-end that learns expert-specific relative utility from counterfactual changes in downstream detection loss. Four feature-space experts operate on the fused P3 representation, and a cost-adjusted Top-2 router selects expert residuals relative to an explicit Identity action. Across five seeds on the scene-disjoint DroneVehicle test split, Utility Top-2 achieves 52.473±0.807 average precision at an intersection-over-union threshold of 0.50 (AP50) and 42.175±0.238 AP50–95 (mean ± standard deviation across seeds), improving AP50–95 by 3.085 points over dual-stream fusion and 2.959 points over four experts always-on. The router activates 1.450 experts per frame, reaches 95.091% Top-2 single-expert Oracle Coverage and 94.836% No-Harm Rate, and processes 84.746 frames per second at 11.800 ms latency. This reduces latency by 11.8% relative to always-on execution. Removing utility supervision reduces AP50–95 by 2.100 points. Utility routing also reaches 43.057 AP50–95 with YOLO11s-OBB and maintains 35.900 and 33.000 AP50–95 under severity-5 single and compound degradations. In a separate tracking extension on the Multi-Drone Multi-Modality dataset for Multi-Object Tracking (M3OT), temporal routing raises higher-order tracking accuracy (HOTA) to 65.958 with ByteTrack and 65.050 with OC-SORT. These results support task-grounded routing for accurate, degradation-adaptive RGB-T detection with lower latency than permanent expert execution under the evaluated profiling protocol.

Authors

Institutions

Publication Details

Journal
Sensors
Published
2026-11-18
DOI
https://doi.org/10.3390/s26227293
Primary Topic
Advanced Neural Network Applications
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

UtilityRoute: Task-Aware Expert Routing for Aerial RGB-T Perception

Jiachang Li, Yinyin Yu, Yong Wu, Zhichao Xin et al.
5 links
Sensors
Advanced Neural Network Applications
article

UtilityRoute: Task-Aware Expert Routing for Aerial RGB-T Perception

Jiachang Li, Yinyin Yu, Yong Wu, Zhichao Xin, Jianping Xing, Yuxuan Xing
article en

Abstract

Unmanned aerial vehicles (UAVs), as dynamic sensing platforms operating under continuously changing observation conditions, benefit from complementary red-green-blue (RGB) and thermal (T) information for object detection. The detection gains of different feature interventions vary across observations, yet existing methods struggle to assess them accurately, potentially leading to ineffective interventions and redundant computation. We present UtilityRoute, an RGB-T front-end that learns expert-specific relative utility from counterfactual changes in downstream detection loss. Four feature-space experts operate on the fused P3 representation, and a cost-adjusted Top-2 router selects expert residuals relative to an explicit Identity action. Across five seeds on the scene-disjoint DroneVehicle test split, Utility Top-2 achieves 52.473±0.807 average precision at an intersection-over-union threshold of 0.50 (AP50) and 42.175±0.238 AP50–95 (mean ± standard deviation across seeds), improving AP50–95 by 3.085 points over dual-stream fusion and 2.959 points over four experts always-on. The router activates 1.450 experts per frame, reaches 95.091% Top-2 single-expert Oracle Coverage and 94.836% No-Harm Rate, and processes 84.746 frames per second at 11.800 ms latency. This reduces latency by 11.8% relative to always-on execution. Removing utility supervision reduces AP50–95 by 2.100 points. Utility routing also reaches 43.057 AP50–95 with YOLO11s-OBB and maintains 35.900 and 33.000 AP50–95 under severity-5 single and compound degradations. In a separate tracking extension on the Multi-Drone Multi-Modality dataset for Multi-Object Tracking (M3OT), temporal routing raises higher-order tracking accuracy (HOTA) to 65.958 with ByteTrack and 65.050 with OC-SORT. These results support task-grounded routing for accurate, degradation-adaptive RGB-T detection with lower latency than permanent expert execution under the evaluated profiling protocol.

SensorsVol. 26(22)
Shandong University (CN)
Ministry of Natural Resources of the People's Republic of China
Sustainable cities and communities
Advanced Neural Network Applications
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon