ParC-Net: Position Aware Circular Convolution with Merits from ConvNets and Transformer

Recently, vision transformers started to show impressive results which outperform large convolution based models significantly. However, in the area of small for mobile or resource constrained devices, ConvNet still has its own advantages both performance and model complexity. We propose ParC-Net, a pure backbone that further strengthens these by fusing merits into ConvNets. Specifically, we p osition ware ci r cular c onvolution (ParC), light-weight op boasts global receptive field while producing location sensitive features as local convolutions. combine ParCs squeeze-excitation ops form meta-former like block, attention mechanism transformers. The aforementioned block can be used plug-and-play manner replace relevant blocks ConvNets Experiment proposed ParC-Net achieves better than popular transformer common tasks datasets, having fewer parameters faster inference speed. For classification on ImageNet-1k, 78.6% top-1 accuracy with about 5.0 million parameters, saving 11% 13% computational cost but gaining 0.2% higher 23% speed (on ARM Rockchip RK3288) compared MobileViT, uses only 0.5 $$\times $$ 2.7% DeIT. On MS-COCO object detection PASCAL VOC segmentation tasks, also shows performance. Source code is available at https://github.com/hkzhang91/ParC-Net .