MSWiM 2025

Title: Small Object Detection in UAV Imagery: Challenges and Solutions

Noura Aljeri,
University of Ottawa, Canada and Kuwait University, Kuwait

Abstract: Unmanned Aerial Vehicles (UAVs) are becoming essential visual monitoring technologies in various fields including urban planning, environmental monitoring, traffic surveillance, and disaster response. Nevertheless, there are some challenges in detecting objects in UAV imagery since the targets can be very small, densely packed, and acquired under unstable conditions such as shifting lighting, cluttered backgrounds, and flying motion. These factors make the already hard job of small object detection (SOD) even harder. SOD needs particular labelling strategies, model modifications, and deployment considerations.
With a focus on small objects, this tutorial guides participants through the whole UAV imagery labelling and object identification pipeline. We start with annotation guidelines and dataset characteristics unique to aerial perspectives, followed by a general overview of the evolution of object detection architectures from two-stage detectors to real-time one-stage frameworks. The focus then moves to the latest YOLO-based techniques, emphasizing attention mechanisms, multiscale feature fusion, architectural modifications, and efficiency-focused designs specifically suited for UAV applications. We will also discuss enhancements to training, improvements in inference time, and accuracy vs. computational efficiency trade-offs for platforms with limited resources.
By the end of the tutorial, participants will have a better understanding of the difficulties associated with UAV-based small objects detection, practical knowledge of labelling and model construction, and a taxonomy of efficient techniques for enhancing deployment in the real world.

Short Bio: Noura Aljeri [Senior Member, IEEE, Member, ACM] is currently an Assistant Professor with the Department of Computer Science at Kuwait University. She received her Ph.D. in Computer Science from the University of Ottawa, Canada, where she was awarded the Most Outstanding PhD Thesis Award (2020) and the Pierre Laberge Award (2021) for her contributions to mobility management in autonomous and connected vehicular networks. Her current research interests span smart mobility, topology management, and prediction models for connected and autonomous vehicular networks, as well as smart transportation systems and the Internet of Drones. More recently, she has been working on UAV imagery analysis, focusing on annotation strategies, small object detection, and real-time deep learning models for aerial surveillance applications. She has published extensively in these areas. Dr. Aljeri has served on the technical program committees of several ACM and IEEE flagship conferences, including Globecom, MSWiM, PE-WASUN, and ICC. She has also served on the Editorial Board of ACM ICPS and is currently on the editorial board of ACM Computing Surveys.

Title: From Perception to Action: A Hands-On Tutorial on Control-Feedback-Driven Adaptive AI Agents in Path-Planning Simulation Environments

Peng Sun,
Duke Kunshan University, China

Abstract: Autonomous systems such as self-driving cars, warehouse robots, and mobile delivery drones must perceive their surroundings accurately while simultaneously making fast and reliable control decisions. These two processes—perception(extracting features from sensor inputs) and control (choosing actions to achieve a goal)—are traditionally designed as separate modules. In many real-world deployments, however, this separation creates significant vulnerabilities: inaccurate perception reduces control quality, while suboptimal control choices can in turn degrade the usefulness of perception.
In dynamic, non-stationary environments—for example, when previously unseen obstacles appear during a robot’s navigation task—static perception or control modules often fail to adapt quickly enough. This leads to slower learning, reduced robustness, or even unsafe behavior. To address these challenges, our recent research proposes a simulation framework for coordinated perception and control coupled with a control-feedback-driven perception optimization mechanism.
This tutorial will introduce these concepts in a practical, hands-on manner, demonstrating how to design, implement, and test AI agents that dynamically co-evolve their perception and control modules. The tutorial is particularly relevant for researchers and practitioners in robotics, reinforcement learning, computer vision for embodied AI, and simulation-based algorithm prototyping.
In this tutorial, participants will learn: (1) How to integrate perception and control models within a unified simulation platform to study their interaction; (2) How to implement a closed-loop feedback mechanism in which control performance metrics guide online updates to the perception model; (3) How to benchmark agents under both static and dynamically changing environments, analyzing robustness, adaptability, and convergence efficiency; and (4) Best practices for simulation-driven research to shorten prototyping cycles before deploying algorithms to real robots.