Enhancing Object Detection Model with Post-Inference Feed-back Mechanism Using Temporal Information

Object detection is essential in applications such as healthcare (e.g. organ detection in CT and ultrasound scans), autonomous driving, and surveillance. However, smaller models like YOLOv8 nano or small often face a drop in accuracy (mAP) in challenging scenarios, such as occlusions, poor lighting, or small object sizes. While larger YOLOv8 models provide better accuracy, they are unsuitable for real-time use due to their size and computational demands. To address these limitations, we propose a novel technique that integrates temporal information through a feedback mechanism, consisting of focused mask generation and box expansion. Focused mask generation uses previous frame detections to enhance inference, while box expansion ensures better object coverage in subsequent frames. Our approach can be applied to any detection model, including those used in medical imaging or surveillance. Evaluations on benchmark datasets and real-world scenarios show a 4 percent improvement in mAP over the baseline YOLOv8 model, demonstrating enhanced detection accuracy and robustness under diverse conditions without compromising real-time performance.