Divide and Conquer: Object Co-occurrence Helps Mitigate Simplicity Bias in OOD Detection

This repository provides the official implementation of the paper:

Divide and Conquer: Object Co-occurrence Helps Mitigate Simplicity Bias in OOD Detection (CVPR 2026)

OCO is an object-centric OOD detection method that leverages Object CO-occurrence patterns and a divide-and-conquer scoring strategy to mitigate simplicity bias and improve both challenging and full-spectrum OOD detection. Below is the pipeline:

Installation

Our implementation is based on the OpenOOD framework. For more detailed environment configuration and instructions, please refer to https://github.com/Jingkang50/OpenOOD.

# Clone the repository
git clone https://github.com/Michael-McQueen/OCO.git
cd OCO

# Create and activate a new Conda environment
conda create -n oco python=3.10
conda activate oco

# Install required dependencies
pip install git+https://github.com/Jingkang50/OpenOOD
pip install libmr
pip install timm

Dataset Preparation

Download the required benchmark datasets using the provided script:

bash scripts/download/download.sh

Note: For the ImageNet-1K training images, due to copyright restrictions, please ensure you have downloaded them from the official ImageNet website and placed them in the appropriate ./data directory structure.

How to Run

Running the OCO pipeline involves a two-step process: pretraining the Slot Attention module, followed by the integrated 3-stage OCO training and evaluation pipeline.

Step 1: Pretrain Slot Attention

Before training the main OCONet, you should obtain a pre-trained Slot Attention checkpoint.

1. Clone and follow the instructions in the DINOSAUR repository to pre-train the model on the default COCO 2017 dataset.
2. Configuration Settings: resize images to 224x224 and set slot_num to 6. Leave other hyperparameters as their defaults.
3. Save the resulting checkpoint.

Step 2: Execute the OCO Pipeline

Once your pretrained checkpoint is in place, you can execute the full OCO methodology on ImageNet-1K using the provided bash script.

bash scripts/ood/oco/oco.sh

The oco.sh script automates the full 3-stage methodology:

• Stage A: Using ID data to train OCONet.
• Stage B: Building $F_{train}$.
• Stage C: OOD/FS-OOD evaluation.

Acknowledgements

This codebase is built upon OpenOOD. We thank the authors for their excellent work.

Citation

If you find this work useful, please cite our paper.