Simple Fusion of Object Detectors for Improved Performance and Faster Deployment

Abstract

Object detectors often suffer from multiple performance limitations which may be attenuated with larger training datasets, improved training techniques, and complex detection models. However, such strategies are complex and time-consuming for applications requiring fast deployments. We propose a Simple Fusion of Object Detectors (SFOD) late ensemble method to combine existing pre-trained, off-the-shelf, fine-tuned object detectors and leverage on their divergences to improve the overall detection performance. Comprehensive experimental evaluations, based on PASCAL VOC07 challenge, demonstrate SFOD’s ability to improve mean average precision ( ${mAP}$ ) for different fusion sizes and base detector combinations, reaching an absolute 84.08% ${mAP}$ and an improvement of 3.97% ${mAP}$ . The improvements extend to most classes, fusion sizes, and base detector combinations, revealing $AP$ improvements up to 17.35% over baselines, for particular object classes. Practical application evaluations, based on optimal threshold selection, also reveal improvements of 10.54% and 8.36% of mean recall ( $mR$ ) and ${mAP}$ , respectively. Our approach does not require additional training and is quickly deployable, yet providing a few adjustable hyperparameters to optimize the recall-precision relation for specific applications. Improvements obtained from our proposed SFOD fusion pipeline span across a broad range of object classes and are important for a wide variety of critical applications where every successful detection is treasured.

Keywords

• Deep learning
• ensemble
• fast deployment
• fusion
• object detection