YOLOE Data Engine Pipeline

This repository contains working scripts for preparing and refining YOLOE grounding and detection training data. The pipeline is currently script-driven and expects dataset paths, text embeddings, model weights, and buffer directories to be configured in the relevant Python or shell script before running.

Pipeline

1. Load grounding or detection labels from JSON/cache files.
   • DataEngine.load_cached_label(...) loads Ultralytics .cache files for grounding or detection data.
   • DataEngineAgent.multi_process_load_grounding_data(...) converts grounding JSON annotations into per-image sample JSON files.
2. Run YOLOE predictions.
   • DataEngine.load_yoloe() loads the configured YOLOE model.
   • DataEngineAgent.multi_process_batch_model_predict(...) writes per-image model prediction JSON files under the configured buffer directory.
3. Merge model predictions into labels.
   • Predictions with high overlap against existing boxes are skipped.
   • Remaining predictions are added to sample labels and saved for cache generation or inspection.
4. Inspect and refine outputs.
   • Visualization helpers such as visual_json.py, data_visual_flickr.py, data_visual_mixed.py, and grounding_dataset_visualizer.py can render generated labels and predictions.
   • refine_text.py refines grounding text prompts and writes updated cache files.

Example Scripts

• do_flickr.sh runs the Flickr refinement and visualization flow.
• do_mixed.sh runs the mixed-grounding refinement flow.
• Edit the hard-coded dataset, model, cache, and environment paths in these scripts before running them.

Notes

• Several scripts assume an Ultralytics checkout and local dataset layout under /root/ultra_louis_work/....
• Use a Python environment with ultralytics, PyTorch, NumPy, Pillow, Matplotlib, and the other imports required by the selected script.
• Generated buffer directories and cache files are local artifacts and are not committed here.