AutFpn

AutFPN stands for Automatic Feature Pyramid Network, a family of methods in computer vision that aim to automatically design the structure of Feature Pyramid Networks used for multi-scale feature representation in tasks such as object detection and instance segmentation. By exploring a search space of architectural choices for how features are propagated, fused, and scaled across pyramid levels, AutFPN seeks to improve accuracy and efficiency compared to manually designed FPNs.

Typical search space in AutFPN includes decisions about whether to include extra pyramid levels (such as P6 and P7), the topology of the top-down and lateral connections, the fusion operation (for example addition or concatenation followed by a 1x1 convolution), channel alignment, and whether to apply attention or weighting mechanisms. Search methods commonly employ differentiable architecture search (DARTS-like) or reinforcement-learning-based controllers, often optimizing a joint objective that balances accuracy with latency or FLOPs. After the search, the best-performing architecture is instantiated and trained with a standard backbone (for example ResNet or EfficientNet) on benchmarks used for detection and segmentation.

AutFPN aims to reduce manual design effort and tailor the pyramid to the data and task at hand, but the approach can be computationally intensive and sensitive to the chosen search space design and training protocol. It is related to conventional Feature Pyramid Networks but emphasizes automated optimization over fixed heuristics. See also: Feature Pyramid Network, Neural Architecture Search.