Toward Efficient Small-Object Detection: Reconstruction Guided Refinement with Adaptive Inference
Faculty Mentor
Min Gyu Kim
Location
Russell Union Room 2080
If Other was choses above, please indicate your topic area here:
Computer Vision & Robotics
Type of Research
On-going
Session Format
Oral Presentation
College
Allen E. Paulson College of Engineering & Computing
Department
Electrical and Computer Engineering
Abstract
Small-object detection in imagery is inherently challenging due to low-resolution object signatures, dense occlusions, and significant background clutter. Although modern detectors can achieve strong average precision, uniformly applying high-resolution inference to every frame imposes prohibitive computational costs for lightweight systems. To address this trade-off between accuracy and efficiency, our work proposes an adaptive detection framework for efficient small-object detection. The framework first performs global detection using a baseline detector enhanced with Cascaded Deformable Dual-Path Attention (CD-DPA), which improves feature representation in cluttered scenes by strengthening both local boundary cues and contextual semantics, enabling more reliable interpretation of object relationships and surrounding background structures. It then evaluates prediction reliability from detector outputs and activates image slicing only for low-confidence cases where global inference is likely to miss small objects. When slicing is triggered, a lightweight self-supervised reconstructor generates a reconstruction-based difference map to localize challenging regions and select a small set of high-value tiles for Slicing-Aided Hyper Inference (SAHI), thereby avoiding exhaustive full-grid slicing. Experiments on the VisDrone2019-DET benchmark demonstrate that the proposed method improves small-object detection performance while reducing overall computational overhead.
Program Description
.
Start Date
4-23-2026 2:00 PM
End Date
4-23-2026 2:15 PM
Recommended Citation
Alam, Md Mahinur and Kim, Min Gyu, "Toward Efficient Small-Object Detection: Reconstruction Guided Refinement with Adaptive Inference" (2026). GS4 Student Scholars Symposium. 137.
https://digitalcommons.georgiasouthern.edu/research_symposium/2026/2026/137
Toward Efficient Small-Object Detection: Reconstruction Guided Refinement with Adaptive Inference
Russell Union Room 2080
Small-object detection in imagery is inherently challenging due to low-resolution object signatures, dense occlusions, and significant background clutter. Although modern detectors can achieve strong average precision, uniformly applying high-resolution inference to every frame imposes prohibitive computational costs for lightweight systems. To address this trade-off between accuracy and efficiency, our work proposes an adaptive detection framework for efficient small-object detection. The framework first performs global detection using a baseline detector enhanced with Cascaded Deformable Dual-Path Attention (CD-DPA), which improves feature representation in cluttered scenes by strengthening both local boundary cues and contextual semantics, enabling more reliable interpretation of object relationships and surrounding background structures. It then evaluates prediction reliability from detector outputs and activates image slicing only for low-confidence cases where global inference is likely to miss small objects. When slicing is triggered, a lightweight self-supervised reconstructor generates a reconstruction-based difference map to localize challenging regions and select a small set of high-value tiles for Slicing-Aided Hyper Inference (SAHI), thereby avoiding exhaustive full-grid slicing. Experiments on the VisDrone2019-DET benchmark demonstrate that the proposed method improves small-object detection performance while reducing overall computational overhead.
