An Empirical Study of Adder Neural Networks for Object Detection
This work addresses energy efficiency for object detection in applications like autonomous driving and face detection, but it is incremental as it adapts an existing method (AdderNets) to a new task.
The paper tackles the problem of reducing energy consumption in object detection by applying Adder Neural Networks (AdderNets), which use addition operations instead of multiplications, and achieves a 37.8% AP on COCO with about 1.4x energy reduction compared to convolutional counterparts.
Adder neural networks (AdderNets) have shown impressive performance on image classification with only addition operations, which are more energy efficient than traditional convolutional neural networks built with multiplications. Compared with classification, there is a strong demand on reducing the energy consumption of modern object detectors via AdderNets for real-world applications such as autonomous driving and face detection. In this paper, we present an empirical study of AdderNets for object detection. We first reveal that the batch normalization statistics in the pre-trained adder backbone should not be frozen, since the relatively large feature variance of AdderNets. Moreover, we insert more shortcut connections in the neck part and design a new feature fusion architecture for avoiding the sparse features of adder layers. We present extensive ablation studies to explore several design choices of adder detectors. Comparisons with state-of-the-arts are conducted on COCO and PASCAL VOC benchmarks. Specifically, the proposed Adder FCOS achieves a 37.8\% AP on the COCO val set, demonstrating comparable performance to that of the convolutional counterpart with an about $1.4\times$ energy reduction.