Beyond MACs: Hardware Efficient Architecture Design for Vision Backbones

The paper introduces LowFormer, a new family of vision backbone networks designed to enhance efficiency in computer vision tasks. It critiques the reliance on MACs as a performance metric, demonstrating that LowFormer, which incorporates a lightweight design feature called Lowtention, achieves superior execution speed and accuracy on benchmarks like ImageNet. This architecture is particularly beneficial for practitioners working on edge devices, as it offers significant speed improvements across various hardware platforms and is adaptable for multiple downstream applications such as object detection and semantic segmentation.