Unofficial GroundingDINO adds training the official repo skipped

What it does This is an independent re-implementation of the Grounding DINO vision-language detector. While the official release only ships inference weights, this repo lets you fine-tune on custom object-detection or visual-grounding data, or pre-train from scratch. It keeps the original model architecture intact, so existing checkpoints remain compatible.

The interesting bit The authors created a custom odvg data format that mixes traditional object-detection annotations and visual-grounding captions in one training pipeline. They also report nudging COCO zero-shot mAP to 55.1 after pre-training on 1.8 million mixed images, and 57.3 after fine-tuning—though the README leaves the exact “training acceleration strategy” unexplained.

Key highlights

• Adds training and fine-tuning capabilities missing from the official GroundingDINO release.
• Supports mixed OD/VG datasets via a custom odvg format, with converters for COCO and GRIT.
• Includes multi-node Slurm training scripts out of the box.
• Retains the original architecture, so inference can reuse the official library and existing checkpoints.
• Published checkpoints include a COCO-fine-tuned model and a 1.8M-image pre-trained model.

Caveats

• Inference still requires installing the official GroundingDINO library; this repo is primarily a training environment.
• The README touts a “training acceleration strategy” but never details what it is.
• Dataset conversion to odvg format is mandatory before training begins.

Verdict Worth a look if you need to fine-tune or pre-train an open-vocabulary detector and have outgrown the official repo’s inference-only weights. Skip it if you just want zero-shot inference without the data-wrangling overhead.