Benchmarking and Mitigating UAV Perception Failure Under Rapid Motion Blur, Scale Shifts, and Occlusion

The problem

Current perception models deployed on unmanned aerial vehicles achieve low absolute accuracy on standard downstream tasks (e.g., recall under 7.5% in scene understanding, under 40% in action recognition, and under 30% m_vIoU in tracking). These failures are driven by real-world flight conditions—rapid camera motion blur, extreme resolution and altitude shifts, severe occlusion, and small target sizes—that standard benchmarks do not systematically couple. As a consequence, aerial vision pipelines remain brittle and cannot be reliably trusted for autonomous navigation, search-and-rescue, or monitoring in complex flight regimes.

Why it matters

Autonomous UAV systems can maintain reliable perception and tracking across agile flight maneuvers and dynamic altitudes without catastrophic performance collapse.

Ways to approach it

Prior-work checks are free with an account. Results someone already ran are shown to everyone.

1. 1

   Systematic Robustness Evaluation: Collect or synthesize a benchmark across existing UAV datasets (e.g., UAV123, DroneAction, VisDrone) that isolates and quantifies degradation curves across calibrated levels of motion blur, altitude-driven resolution scaling, and occlusion rates, measuring task performance drops against clean baselines.

   Check for prior work
2. 2

   Motion- and Scale-Aware Pre-Restoration Adapters: Implement lightweight, task-coupled front-end restoration and scale-adaptation modules conditioned on UAV flight dynamics (IMU/velocity priors or optical flow magnitude) and measure end-to-end downstream mAP, recall, and tracking IoU.

   Check for prior work
3. 3

   Multi-Scale Temporal Aggregation: Develop a temporal feature aggregation pipeline that leverages multi-frame redundancy across video streams to recover degraded small objects and blurred visual cues, evaluating performance recovery under aggressive UAV maneuver sequences.

   Check for prior work

Why it might fail

If low-level image deblurring and super-resolution front-ends can be trivially plugged into existing off-the-shelf vision models to close the performance gap without requiring aerial-specific domain modeling.

Evidence

Each paper's own statement of the limitation, verbatim.

• Where Does It Exist from the Low-Altitude: Spatial Aerial Video GroundingNeurIPS 2025

  Performance is modest in absolute terms (m_vIoU 27.15%), indicating the task remains largely unsolved for small objects, occlusion, and rapid UAV motion

• 4D Panoptic Scene Graph GenerationNeurIPS 2023

  Absolute performance is very low (all R@100 below 7.5, mR@100 below 5), so the task is largely unsolved and the model handles only simple scenes, failing in complex real-world environments

• UAV-Human: A Large Benchmark for Human Behavior Understanding With Unmanned Aerial VehiclesCVPR 2021

  Absolute performance on most tasks remains low (e.g., <40% for action recognition), indicating that current models struggle with extreme UAV motion blur and resolution shifts

Nearest existing work

• SPUR: Scale-Partitioned Uncertainty Rectification for Robust UAV-on-UAV InterceptionICML 2026
• AeroGS: Scale-Aware Gaussian Splatting for Pose-Free Dynamic UAV Scene ReconstructionCVPR 2026
• Breaking Smooth-Motion Assumptions: A UAV Benchmark for Multi-Object Tracking in Complex and Adverse ConditionsCVPR 2026
• The Unmanned Aerial Vehicle Benchmark: Object Detection and TrackingECCV 2018
• No Labels, No Look-Ahead: Unsupervised Online Video Stabilization with Classical PriorsCVPR 2026
• Adaptive 3D Perception for Small Aerial Targets Under Sparse Sampling via Reinforcement LearningCVPR 2026
• Resource-Efficient RGBD Aerial TrackingCVPR 2023
• PiLoT: Neural Pixel-to-3D Registration for UAV-based Ego and Target Geo-localizationCVPR 2026
• Adaptive Sparse Convolutional Networks With Global Context Enhancement for Faster Object Detection on Drone ImagesCVPR 2023
• Fusion Meets Diverse Conditions: A High-diversity Benchmark and Baseline for UAV-based Multimodal Object Detection with Condition CuesICCV 2025
• UAV-CB: A Complex-Background RGB-T Dataset and Local Frequency Bridge Network for UAV DetectionCVPR 2026
• UAVLight: A Benchmark for Illumination-Robust 3D Reconstruction in Unmanned Aerial Vehicle (UAV) ScenesCVPR 2026
• Cross-modal Fuzzy Alignment Network for Text-Aerial Person Retrieval and A Large-scale BenchmarkCVPR 2026
• AeroDGS: Physically Consistent Dynamic Gaussian Splatting for Single-Sequence Aerial 4D ReconstructionCVPR 2026
• UAST: Unified Active Search and Tracking for Arbitrary Targets with UAVsCVPR 2026

Related open problems