Visual recognition models can be made cheap to run at inference time along four largely independent axes: designing light-weight static backbones (split-transform-merge blocks, inverted bottlenecks, feature reuse, efficient self-attention) or searching them automatically with NAS; making the computational graph dynamic so that easy inputs, uninformative image regions and redundant video frames receive less computation; specialising detection and segmentation heads for real-time use; and compressing a trained network by pruning, quantization, low-rank factorization or knowledge distillation. A student learns to reason about the accuracy-versus-cost trade-off, to choose among these families for a given deployment target, and to distinguish theoretical FLOP savings from realised latency on actual hardware.
This survey by Wang, Han, Wang, Song, Tian, and Huang (published August 2023) addresses computational challenges in computer vision systems. It examines how to balance model effectiveness with comput…