Robust Curriculum Selection
Temporal statistics of sample losses separate hard-but-informative images from ambiguous or noisy samples, forming a stable learning route.
2
Learning phases that first stabilize optimization and then deliberately increase task difficulty.
0
Extra inference parameters introduced by the training framework.
CECS
Designed for context-entangled content segmentation, including camouflaged scenes.
Method
Stabilize, then challenge.
CurriSeg models learning as a progression: robust curriculum selection first builds reliable representations, then anti-curriculum promotion forces the network to discover deeper contextual cues.
TSSW and PUE
Temporal sample weighting and pixel-level uncertainty estimation reduce unstable early gradients while keeping useful difficulty signals.
Anti-Curriculum Promotion
Spectral-Blindness Fine-Tuning suppresses high-frequency shortcuts, encouraging low-frequency structural and contextual reasoning.
Framework
Dual-phase training for context-entangled scenes.
The framework combines temporal difficulty modeling, uncertainty-aware weighting, and spectral perturbation in one training pipeline.
Visual Comparison
Sharper masks across CECS scenarios.
CurriSeg improves baselines over camouflaged object detection, polyp image segmentation, transparent object detection, and concealed defect detection examples.
Usage
Research code layout.
The repository provides PyTorch training, anti-curriculum fine-tuning, and testing scripts. Common CECS/COD benchmarks can be found in awesome-concealed-object-segmentation.
Train
python Train.py \
--train_root YOUR_TRAININGSETPATH \
--val_root YOUR_VALIDATIONSETPATH \
--save_path YOUR_CHECKPOINTPATH
Promote
python anti_curri_stage.py \
--train_root YOUR_TRAININGSETPATH \
--val_root YOUR_VALIDATIONSETPATH \
--save_path YOUR_ANTI_CURRI_CHECKPOINTPATH \
--load YOUR_CHECKPOINTPATH/Net_epoch_best.pth \
--use_sbft
Test
python Test.py \
--pth_path YOUR_CHECKPOINTPATH/Net_epoch_best.pth \
--test_dataset_path YOUR_TESTINGSETPATHContact
Questions and collaboration.
Please contact us via email at chunminghe19990224@gmail.com or chunming.he@duke.edu.
Citation
Cite CurriSeg.
@inproceedings{he2026curriseg,
title={Refining Context-Entangled Content Segmentation via Curriculum Selection and Anti-Curriculum Promotion},
author={He, Chunming and Zhang, Rihan and Xiao, Fengyang and Zhang, Dingming and Cao, Zhiwen and Farsiu, Sina},
booktitle={International Conference on Machine Learning (ICML)},
year={2026}
}