DIS (IS-Net)

Dichotomous Image Segmentation with intermediate supervision strategy

Dichotomous Image Segmentation — proposes IS-Net with 3 components:

1. Ground truth (GT) encoder
2. Image segmentation component ($U^2$-Net based)
3. Intermediate supervision strategy

Stage 1: Self-supervised GT Encoder Training

$L_{gt} = \sum_{d=1}^{D} \lambda_{d}^{gt} BCE(F_{gt}(\theta_{gt}, G)_d, G)$

GT encoder is frozen after this stage.

Stage 2: Feature Consistency

Feature Consistency Loss (intermediate supervision):

$L_{fs} = \sum_{d=1}^{D} \lambda_{d}^{fs} |f_{d}^{I} - f_{d}^{G}|^2$

$L_{sg} = \sum_{d=1}^{D} \lambda_{d}^{sg} BCE(F_{sg}(\theta_{sg}, I), G)$

Total loss: $L = L_{fs} + L_{sg}$

Results

Metric: Human Correction Efforts (HCE)