DRNet

DRNet enhances few-shot semantic segmentation by using self-adapted and cross-attended recalibration modules to improve query-image segmentation from a limited number of annotated support images and to mitigate high intra-class variance between query and support instances.

Key Features:

• Self-adapted Recalibration (SR) Module: Leverages semantic-aware information from the query image and uses support-set guidance to predict an initial (incomplete but correct) object region, then extracts feature embeddings from that region to refine segmentation.
• Cross-attended Recalibration (CR) Module: Refines each pixel response in the query feature map by attending to all foreground pixels in the support feature map, computing a similarity-weighted average and integrating it via a residual connection to propagate semantic knowledge from support to query.
• Performance and Validation: Evaluated on PASCAL-5^i and COCO-20^i benchmarks, achieves mean IoU of 63.6% (1-shot) and 64.9% (5-shot) on PASCAL-5^i, and 44.7% (1-shot) and 49.6% (5-shot) on COCO-20^i, surpassing existing state-of-the-art methods reported on these benchmarks.

Scientific Applications:

• Medical imaging: Enables segmentation with limited annotated examples where intra-class variance across instances is high.
• Remote sensing: Supports object delineation in satellite and aerial imagery from a small number of labeled examples.
• Precise object delineation with limited labels: Applicable to any field requiring accurate object boundaries when labeled examples are scarce.

Methodology:

DRNet combines a Self-adapted Recalibration (SR) module and a Cross-attended Recalibration (CR) module that jointly refine and propagate support-derived semantic features to mine more accurate target regions in query images.