Refining Context-Entangled Content Segmentation via Curriculum Selection and Anti-Curriculum Promotion

TL;DR

Context-Entangled Content Segmentation (CECS) is challenging because objects share intrinsic patterns with their surroundings, leading to brittle learned representations. CurriSeg introduces a two-phase learning framework: Robust Curriculum Selection (RCS) to stabilize optimization via temporal loss statistics and pixel-level uncertainty, and Spectral-Blindness Fine-Tuning (SBFT) to induce reliance on low-frequency structural cues by attenuating high-frequency information. The approach yields consistent performance gains across camouflaged object detection, polyp segmentation, transparent object detection, and defect segmentation, without adding parameters or increasing training time, and it generalizes to weak/semi-supervised, multi-modality, and video segmentation settings. This work provides a principled, architecture-agnostic training paradigm for robust, context-aware dense prediction under entangled data distributions.

Abstract

Biological learning proceeds from easy to difficult tasks, gradually reinforcing perception and robustness. Inspired by this principle, we address Context-Entangled Content Segmentation (CECS), a challenging setting where objects share intrinsic visual patterns with their surroundings, as in camouflaged object detection. Conventional segmentation networks predominantly rely on architectural enhancements but often ignore the learning dynamics that govern robustness under entangled data distributions. We introduce CurriSeg, a dual-phase learning framework that unifies curriculum and anti-curriculum principles to improve representation reliability. In the Curriculum Selection phase, CurriSeg dynamically selects training data based on the temporal statistics of sample losses, distinguishing hard-but-informative samples from noisy or ambiguous ones, thus enabling stable capability enhancement. In the Anti-Curriculum Promotion phase, we design Spectral-Blindness Fine-Tuning, which suppresses high-frequency components to enforce dependence on low-frequency structural and contextual cues and thus strengthens generalization. Extensive experiments demonstrate that CurriSeg achieves consistent improvements across diverse CECS benchmarks without adding parameters or increasing total training time, offering a principled view of how progression and challenge interplay to foster robust and context-aware segmentation. Code will be released.

Figures (8)

• Figure 1: Performance on CECS data. Top: CL is short for curriculum learning bengio2009curriculum. Concealed object masks are highlighted in purple and blue, overlaid on the original data for visual clarity. Bottom: "Task (Dataset, Metric)". The radar chart shows that training the baseline under our CurriSeg brings performance gains than the standard training manner, where TOP-1 is the composite baselines with the top metric scores per task.
• Figure 2: Framework of CurriSeg. In (1), performance improves steadily; brighter regions in $\omega \times {W}$ indicate more trustworthy pixels. In (2), Phase I applies Robust Curriculum Selection with sample/pixel-level weighting, while Phase II includes Anti-Curriculum Promotion via spectral filtering. TSD describes how to distinguish hard-but-informative samples from ambiguous or outlier ones.
• Figure 3: Visual comparison on COD, PIS, TOD, and CDD tasks.
• Figure 4: Sensitivity analysis of hyperparameters, including $K$, $p_{min}$, $\sigma^*$, $\gamma$, $W_{min}^s$, $W_{min}$, $r$, and $T_c$.
• Figure 5: Visualization of attention map and t-SNE of foreground (orange) and background (blue) features. (b)-(d) and (f)-(h) correspond to the prediction, attention map, and t-sne of FEDER trained under the standard setting and our CurriSeg framework.