Advancing object detection models: an investigation focused on small object detection in complex scenes

Abstract

Small object detection remains a persistent challenge in computer vision, especially in safetycritical applications, such as autonomous driving and aerial surveillance, where objects of interest often occupy only a few pixels and are easily lost in cluttered scenes. To advance the performance of small object detection models, this thesis proposes two novel approaches focused on increasing both accuracy and robustness. The first approach introduces a semantic segmentation-guided feature fusion framework, where contextual cues from a segmentation model are integrated into the object detection pipeline. A lightweight attention mechanism is used to merge semantic and visual features, enhancing the detection of small objects. The experimental results demonstrate clear improvements in identifying challenging small targets, proving the effectiveness of cross-task feature integration. The second approach utilizes feature pyramidal structures to improve multi-scale feature representation through a novel dilated strip-wise spatial feature pyramid, which employs dilated stripwise depth convolutions. Evaluated on the VisDrone and AI-TOD benchmark datasets, this model shows significant improvements over the baseline, effectively detecting objects in densely packed environments. The approach achieves state-of-the-art performance on the AI-TOD dataset. Together, these approaches offer distinct strategies for overcoming the limitations of the existing object detection models. The research findings emphasize the importance of both semantic guidance and spatial feature refinement in enhancing small object detection.