This repository provides the official implementation of DitHub, accepted at NeurIPS 2025.

DitHub is a modular and scalable framework designed for Incremental Open-Vocabulary Object Detection. Inspired by version control systems, DitHub enables the seamless management, update, and composition of class-specific adaptation modules.

⚙️ Setup

Prerequisites

• Python >= 3.9
• CUDA >= 11.8, GCC >= 11.4

Installation

• Clone the DitHub repository from GitHub and change the current directory

git clone https://github.com/chiara-cap/DitHub.git
cd DitHub

• Create and activate a conda environment

conda create -n dithub python=3.9 pip
conda activate dithub

• Install the requirements

pip install -r requirements.txt

• Clone the Deformable-DETR repository and compile CUDA operators

git clone https://github.com/fundamentalvision/Deformable-DETR.git

cd Deformable-DETR/models/ops
sh ./make.sh
# unit test (should see all checking is True)
python test.py

• Download the pre-trained GroundingDINO model weights to be used as the starting point for training.

wget -q https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth

📁 Datasets

To reproduce the experiments in the paper, download the following datasets and place them inside the datasets/ folder:

• COCO — used for zero-shot evaluation
• ODinW-35 — required because it includes the 13 subsets used in the paper (ODinW-13) for incremental training

To download ODinW-35, run the following script:

mkdir datasets
cd datasets
python tools/download_odinw.py

ℹ️ Note: ODinW-13 is automatically configured when downloading ODinW-35.

To test on the ODinW-O benchmark (for overlapping classes), after downloading ODinW-35, run:

python tools/overlapped_classes_dataset.py

🚀 Running Dithub

Training

To start the training process, run the following command. This will use the specified configuration, load the pre-trained checkpoint (if available), and save the outputs in the designated directory:

sh train_dithub.sh

Evalutation

Before running evaluation, download the pre-trained DitHub model weights:

wget -q https://github.com/chiara-cap/DitHub/releases/download/v.0.1/model_final.pth

Then, to evaluate the model, run:

sh eval_dithub.sh

Command Line Arguments

Parameter	Description
--config-file	Path to the training configuration file.
--model-config-file	Path to the model architecture configuration file.
--model-checkpoint-path	Path to the pre-trained model checkpoint file (Grounding DINO or DitHub).
--output-dir	Directory where outputs (logs, checkpoints) will be saved.
--dithub	Flag to integrate Dithub into Grounding DINO.
--eval-only	Run the model in evaluation mode without training.

📜 Citation

If you find DitHub useful for your research, please consider citing our paper:

@article{cappellino2025dithub,
   title={DitHub: A Modular Framework for Incremental Open-Vocabulary Object Detection},
   author={Cappellino, Chiara and Mancusi, Gianluca and Mosconi, Matteo and Porrello, Angelo and Calderara, Simone and Cucchiara, Rita},
   journal={Advances in Neural Information Processing Systems},
   year={2025}
 }