panet

PANet, short for Path Aggregation Network, is a convolutional neural network architecture designed to improve feature fusion for instance segmentation. Introduced in 2018 as an extension to the Feature Pyramid Network (FPN), PANet aims to propagate semantic information across multiple feature scales and enhance the quality of object masks.

The core idea of PANet is to build a bottom-up path augmentation on top of the FPN, creating a bidirectional feature pyramid that strengthens information flow from high-level to low-level layers while preserving spatial detail. This bottom-up path runs from the most semantic, high-level feature maps back toward the lower-resolution maps, enabling richer multi-scale representations. In addition, PANet includes a mask refinement module that aggregates features from different pyramid levels when predicting the binary mask for each region of interest. By combining features across scales, the mask head can produce more precise segmentation, particularly for small or crowded objects.

Implementation and usage: PANet is designed for instance segmentation within two-stage detectors such as Mask R-CNN or Faster R-CNN, typically employing ROI Align and a dedicated mask head. It is evaluated on standard benchmarks like COCO, where it has demonstrated improvements in mask average precision (AP) and overall detection performance over FPN-based baselines.

Impact and legacy: PANet has influenced subsequent research on multi-scale feature fusion and refined segmentation heads, serving as a reference design for architectures that seek to enhance cross-scale information flow within detection frameworks. It generally entails higher computational cost than a plain FPN due to the added bottom-up path and mask refinement computations.