This item is available under a Creative Commons License for non-commercial use only
2. ENGINEERING AND TECHNOLOGY
In this paper we investigate a method to reduce the number of computations and associated activations in Convolutional Neural Networks (CNN) by using bitmaps. The bitmaps are used to mask the input images to the network that fall within a rectangular window but do not fall within the boundaries of the objects the network is being trained upon. The mask has the effect of rendering the operations on these portions of the training images trivial. The thesis is that applying this approach to CNNs will not degrade accuracy while at the same time reducing the computational workload and reducing memory footprint. We found that we can remove up to 60% of the input images and see no decrease in accuracy. This leads to activation sparsity that can be exploited using a hardware accelerator to speedup training and inference, and decrease energy consumed.
Warde, A., Yous, H., Gregg, D. & Moloney, D. (2019). No room for squares: using bitmap masks to improve pedestrian detection using CNNs. IMVIP 2019: Irish Machine Vision & Image Processing, Technological University Dublin, Dublin, Ireland, August 28-30. doi: 10.21427/9vge-x813