Data-adaptive binary neural networks for efficient object detection and recognition

Binary neural networks (BNNs) are promising for computational resource-limited devices, but the degradation of feature representation capacity stifles performance due to binarization. The reason is that existing methods fail to adapt to their input when approximating full-precision features. In this paper, we introduce the DA-BNN, a data-adaptive amplitude method based on spatial and channel attention. We generate an adaptive amplitude for a better feature approximation and minimize the gap between the real-valued and 1-bit convolution. Our adaptive amplitude introduces negligible storage but can significantly enhance the performance. Extensive experiments on object detection and recognition are conducted for the comprehensive evaluation of our methods. Our method achieves 64.0% on Pascal VOC with saving of the storage and computation by 18.62× and 15.77×, respectively. While on ImageNet, compared to the full-precision counterpart, 11.04× and 10.80× saving on storage and computation are obtained with just 3% drop on accuracy, demonstrating the effectiveness on both objective detection and recognition tasks.