This repository provides a third-party implementation of the Grounding DINO paper, enabling open-set object detection and grounding. It is designed for researchers and practitioners looking to fine-tune or pre-train models for custom datasets, offering capabilities beyond the official release.

How It Works

The project leverages the DINO architecture, enhanced with grounded pre-training. It supports a custom odvg data format for training, which unifies object detection (OD) and visual grounding (VG) datasets. This approach allows for flexible data integration and training on diverse datasets, including large-scale ones like GRIT-20M.

Quick Start & Requirements

• Install: Clone the repository, install dependencies via pip install -r requirements.txt, and compile custom ops (cd models/GroundingDINO/ops && python setup.py build install).
• Prerequisites: Python 3.7.11, PyTorch 1.11.0, CUDA 11.3.
• Resources: Requires downloading pre-trained models and BERT weights.
• Docs: Dataset conversion details are in data_format.md.

Highlighted Details

• Supports training on both object detection and grounding datasets.
• Implements Slurm multi-machine support and training acceleration strategies.
• Offers inference capabilities with provided pre-trained models.
• Allows fine-tuning and pre-training from scratch.

Maintenance & Community

• The project is actively maintained by its authors, Zuwei Long and Wei Li.
• Contact information for suggestions and bug reports is provided. Contributions via pull requests are welcomed.

Licensing & Compatibility

• The repository does not explicitly state a license in the README.
• Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The project is a third-party implementation and may not perfectly replicate all features or performance of the official Grounding DINO. Training support for object detection data was initially marked as '✖' in the README's feature table, though the text indicates training is supported. Evaluation on custom test sets requires careful configuration of use_coco_eval and label_list.