README.md

March 17, 2026 · View on GitHub

Geometry-driven OOD Detectors Are Class-Incremental Learners (CVPR26)

Abstract

Class-Incremental Learning (CIL) seeks to acquire new classes over time without erasing prior knowledge. While recent methods leverage pre-trained models (PTMs) to curb forgetting, they largely optimize the feature extractor and overlook the crucial classification head. In this work, we advance a simple view: if each task is equipped with a classifier that has the ability to both recognize in-distribution (IND) classes and reject out-of-distribution (OOD) inputs, CIL arises naturally—inputs are accepted only by heads that deem them in-distribution and rejected otherwise. Supported by rigorous theoretical and empirical studies, we find that this ability is characterized by \textbf{Inter-class Separation} and \textbf{Intra-class Compactness}; lacking these, standard linear and cosine-similarity heads remain closed-set and fail to yield a usable OOD signal. To address this, we propose \textbf{GOD} (\textbf{G}eometry-driven \textbf{O}OD \textbf{D}etectors), which unifies IND recognition and OOD rejection in a single geometric space by replacing the learnable head with fixed Equiangular Tight Frame (ETF) anchors; an ETF loss enforces inter-class separation, and an ArcFace loss further tightens intra-class compactness. For efficiency, we further introduce a parameter-efficient hybrid architecture and an efficient inference strategy, thus reducing both parameter footprint and inference cost. Extensive experiments on multiple incremental settings and datasets show that GOD achieves state-of-the-art results.

Run experiment

To get started, set up a conda environment and install the requirements listed by our repo.

conda env create -f environment.yml

Datasets

We have implemented the pre-processing datasets as follows:

  • CIFAR100: will be automatically downloaded by the code.
  • ImageNet-R: Google Drive: link or Onedrive: link
  • ImageNet-A: Google Drive: link or Onedrive: link
  • OmniBenchmark: Google Drive: link or Onedrive: link
  • Car196: link
  • CUB: link

Pre-trained weights

When training , you should specify the folder of your dataset in utils/data.py.

  def download_data(self):
     assert 0,"You should specify the folder of your dataset"
     train_dir = '[DATA-PATH]/train/'
     test_dir = '[DATA-PATH]/val/'

When using the pre-trained weights of iBOT and DINO, you should download the corresponding pre-trained weights to the Moal/checkpoints/ folder.

  def vit_base_patch16_224_adapter_dino(pretrained=False, **kwargs):
      model = VisionTransformer(patch_size=16, embed_dim=768, depth=12, num_heads=12, mlp_ratio=4, qkv_bias=True,
          norm_layer=partial(nn.LayerNorm, eps=1e-6), **kwargs)
      ckpt = torch.load('Your_path', map_location='cpu')

To Train

python main.py --config=exps\[dataset_name].json

Acknowledgments

We thank the following repos providing helpful components/functions in our work.