CARVE – Class-Agnostic Retail Segmentation & Verification Engine

This repository provides the full experimental code and data for the CARVE (Class-Agnostic Retail Segmentation & Verification Engine) paper. It implements the CARVE pipeline that composes GroundingDINO, SAM2, and a CLIP-based verifier for open-vocabulary product discovery, alongside a YOLOv11 segmentation baseline. The release includes scripts for data preprocessing, CLIP cutout generation and fine-tuning, YOLO training, and end-to-end evaluation, enabling both re-running published results and retraining models from scratch.