AViD

Why we included this project

A general open-vocabulary detector often misses objects that matter to a specific team, like a particular product line or a niche part. AViD wraps Grounding DINO in a fine-tuning workflow that adapts it to your own labels without retraining every parameter. It reads bounding-box annotations from CSV files, applies LoRA across the detector's attention and projection layers, and can use EMA to keep training stable. The repo also includes a detection-focused evaluation loop with mAP, precision, recall, and per-class summaries, plus YAML-driven training, single-image inference, and a Gradio demo. For engineers with a small labeled dataset, that is a ready-made path to a domain-specific grounding model instead of assembling the pieces by hand.