NAS-OD: Neural Architecture Search for Object Detection

Recent advancements in object detection have been remarkable, yet the increasing complexity and computational cost of state-of-the-art detectors restrict their use in applications with limited resources. This paper introduces NAS-OD, a novel approach utilizing Neural Architecture Search (NAS) to design a more efficient and lightweight network architecture. NAS-OD employs a hybrid search methodology, combining weight-sharing in the backbone network for enhanced search efficiency and a differentiable strategy to minimize search costs. This innovative approach allows NAS-OD to outperform current state-of-the-art detectors with significantly lower model complexity, evidenced by reduced FLOPS and parameters. Our NAS-OD architecture, optimized on the ImageNet dataset, achieves an impressive accuracy of 71.4% with only 4.8 million parameters. Furthermore, when applied to the PASCAL VOC dataset, NAS-OD attains a mean Average Precision (mAP) of 64.4%, demonstrating its effectiveness in object detection tasks.