---
license: mit
pretty_name: MV-GEL
task_categories:
- image-to-text
- visual-question-answering
tags:
- cad
- multi-view
- geometry
- abc-dataset
size_categories:
- 1K<n<10K
---

# MV-GEL

Multi-view CAD geometric-entity localization dataset derived from the ABC CAD
dataset. Contains 5,903 CAD models, each with corrected multi-view renders,
marked entity views (faces/edges), geometry files, and per-entity QA logs.

## Contents

The CAD data is stored as gzip-compressed tar shards (`cads_<start>_to_<end>.tar.gz`),
each holding ~250 model folders. Extract all shards into a single directory:

```bash
for f in cads_*.tar.gz; do tar -xzf "$f"; done
```

Each model folder (`00000000/`, `00000001/`, …) contains:

- `*.obj`, `*.step` — geometry
- `mesh_views_corrected/` — corrected multi-view renders
- `view_*_marked_*.png` — entity-marked views (faces/edges)
- `views_and_ques_edge_augmented_corrected.log`
- `views_and_ques_face_augmented_corrected.log`

## Splits & mapping

- `train_dataset_central.log` — training model paths
- `val_dataset_central.log` — validation model paths
- `cad_index_mapping.csv` — maps each central id back to its original
  ABC source (`new_id, source_dataset, original_id`)