GhostNet

Family of efficient CNN architectures using cheap linear operations to generate ghost feature maps. GhostNet (CVPR 2020) proposed the Ghost module for efficient feature generation. GhostNetV2 (NeurIPS 2022 Spotlight) added hardware-friendly DFC attention for long-range dependencies, achieving 75.3% top-1 on ImageNet at 167M FLOPs. GhostNetV3 (2024) explored training strategies for compact models, reaching 79.1% top-1 at 269M FLOPs.