GGR: Geometric Gradient Rectification for Safe Open-Set Semi-Supervised Learning

Official implementation of the ECCV 2026 paper "Geometric Gradient Rectification for Safe Open-Set Semi-Supervised Learning".

GGR is a plug-in optimization framework that reformulates open-set SSL as gradient-level control rather than sample-level selection. Given a supervised gradient $g_s$ and an auxiliary gradient $g_u$, GGR rectifies $g_u$ so that the applied auxiliary update is first-order non-opposing with respect to the supervised anchor. Three rectifiers are provided:

surgery_mode	Rectifier	Update rule
vlr (default)	Vector-Level Rectifier	$\tilde g_u = g_u - \dfrac{\langle g_u, g_s\rangle}{\lVert g_s \rVert^2} g_s$ when $\langle g_u, g_s\rangle < 0$, else $\tilde g_u = g_u$
osr	Orthogonal Subspace Rectifier	$\tilde g_u = (I - UU^\top) g_u$
csr	Conic Subspace Rectifier	$\tilde g_u = g_u - U,\min(U^\top g_u, 0)$

Installation

conda create -n ggr python=3.10 -y
conda activate ggr
pip install -r requirements.txt

Dataset Layout

All datasets are expected under ./data:

./data
├── cifar10/cifar-10-batches-py
├── cifar100/cifar-100-python
├── cifar10_openset
├── cifar100_openset
├── imagenet30
└── ood_data

Training

CUDA_VISIBLE_DEVICES=0 python train.py --c config/openset_cv/ggriomatch/ggriomatch_cifar10_6_150_0.yaml

Helper scripts:

bash scripts/train_cifar10.sh 0 ggriomatch
bash scripts/train_cifar100_20.sh 0 ggropenmatch
bash scripts/train_cifar100_50.sh 0 ggrdac
bash scripts/train_in30.sh 0 ggriomatch

Checkpoints go to ./saved_models/openset_cv/<algorithm>/<save_name>/.

Config naming

Configs live under config/openset_cv/<algorithm>/.

• CIFAR-10 / CIFAR-100 base: <algorithm>_<dataset>_<num_classes>_<num_labels>_<seed>.yaml
• ImageNet-30 base: <algorithm>_in30_<percent>_<seed>.yaml with percent in {p1, p5}

Evaluation

Closed-set accuracy on the in-distribution validation split is reported in the training log at every num_eval_iter iterations, and the best checkpoint is saved as model_best.pth. For the open-set evaluation, which measures balanced accuracy across the in-distribution classes and the "unknown" category on the full test set (optionally with auxiliary OOD data), run eval.py against a saved checkpoint:

python eval.py \
  --c config/openset_cv/ggriomatch/ggriomatch_cifar10_6_150_0.yaml \
  --load_path saved_models/openset_cv/ggriomatch/ggriomatch_cifar10_6_150_0/model_best.pth

Acknowledgments

Built on top of:

• USB
• unified_ossl_benchmarks
• MTCF
• OpenMatch
• Safe-Student
• IOMatch
• UAGreg
• DivCon (DAC)

Citation

@misc{chen2026geometricgradientrectificationsafe,
      title={Geometric Gradient Rectification for Safe Open-Set Semi-Supervised Learning}, 
      author={Jiahe Chen and Qian Shao and Qiyuan Chen and Jiaying He and Jintai Chen and Jian Wu and Hongxia Xu},
      year={2026},
      eprint={2606.26973},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.26973}, 
}