Tailoring Backbone Architectures for SSD

This paper introduces an innovative approach to improve Single Shot Detection (SSD) via a targeted channel search strategy. Traditional algorithms often rely on pre-trained weights from classification tasks, employing transfer learning techniques for object detection which do not consistently yield optimal results. Our proposed method enhances SSD performance by specifically optimizing the channel configurations within the backbone network. This strategy ensures effective computational resource utilization while maintaining high levels of detection accuracy. Notably, our approach attains a mean average precision (mAP) of 71.1% on the Pascal VOC dataset, surpassing several established NAS methodologies.