HFA

HFA performs dense semantic transformation between query and support images to establish pixel-wise correlations that improve few-shot semantic segmentation.

Key Features:

• Dense semantic transformation: Performs a dense semantic transformation between query and support images to capture cross-image semantics for segmentation.
• Bilinear feature activation: Uses a bilinear model to establish pixel-wise dense correlations (bilinear feature activation) between query and support features.
• Low-rank decomposition: Applies low-rank decomposition to accelerate bilinear feature activation with negligible performance cost.
• Semantic diffusion: Integrates a semantic diffusion procedure to enhance global harmony and local consistency of activated features.

Scientific Applications:

• Few-shot semantic segmentation: Improves segmentation performance when only limited annotated support images are available, demonstrated on PASCAL VOC and MS COCO.
• Image analysis with limited annotations: Applicable to image-analysis scenarios requiring precise segmentation from minimal training samples.

Methodology:

Dense semantic transformation via a bilinear model to compute pixel-wise dense correlations (bilinear feature activation), accelerated by low-rank decomposition, followed by semantic diffusion to reinforce global harmony and local consistency.