Datasets:

yk3701208
/

random005

@@ -1,208 +0,0 @@
----
-license: cc-by-4.0
-task_categories:
-- image-to-text
-- object-detection
-- visual-question-answering
-size_categories:
-- 1M<n<10M
-language:
-- en
-- zh
-tags:
-- multimodal
-- vision
-- grounding
-- detection
-- ocr
-pretty_name: Locany Multimodal Dataset
----
-# Locany Multimodal Dataset
-## Dataset Description
-This is a large-scale multimodal dataset for vision-language tasks including object detection, grounding, OCR, and UI understanding.
-## Dataset Statistics
-- **Total Unique Images**: 9,860,204
-- **Total Annotations**: 28,768,853
-- **Total Size**: 2867.81 GB
-- **Categories**: 6
-### Breakdown by Category
-| Category | Datasets | Annotations |
-|----------|----------|-------------|
-| Detection | 32 | 16,334,270 |
-| Grounding | 8 | 2,110,597 |
-| Layout | 15 | 1,141,706 |
-| OCR | 13 | 2,060,741 |
-| Pointing | 3 | 876,872 |
-| UI | 16 | 6,244,667 |
-## Dataset Structure
-```
-locany-dataset/
-├── images_hf/              # Parquet files organized by category/subfolder
-│   ├── Detection/
-│   │   ├── COCO/
-│   │   ├── Object365/
-│   │   └── ...
-│   ├── Grounding/
-│   ├── OCR/
-│   ├── UI/
-│   └── ...
-└── annotations_hf/         # JSONL annotation files by category
-    ├── Detection/
-    ├── Grounding/
-    ├── OCR/
-    └── ...
-```
-### Parquet Files (Images)
-Images are stored in Parquet format with two columns:
-- `image`: Raw image bytes (PNG format)
-- `image_path`: Path string (e.g., `images_hf/Detection/COCO/000000178538.jpg`)
-Each parquet file is approximately 5GB.
-### Annotation Files (JSONL)
-Each line in the JSONL files contains:
-```json
-{
-  "image": "images_hf/Detection/COCO/000000178538.jpg",
-  "conversations": [
-    {"from": "human", "value": "<image>\nDetect all objects..."},
-    {"from": "gpt", "value": "<ref>person</ref><box>x1 y1 x2 y2</box>..."}
-  ]
-}
-```
-### Coordinate Format
-Bounding boxes use normalized coordinates in range [0, 1000]:
-- `<box>x1 y1 x2 y2</box>` for rectangles
-- `<box>x y</box>` for points
-To convert to absolute coordinates:
-```python
-absolute_x = normalized_x * image_width / 1000
-absolute_y = normalized_y * image_height / 1000
-```
-## Usage
-### Loading Images from Parquet
-```python
-import pyarrow.parquet as pq
-from PIL import Image
-import io
-# Read parquet file
-table = pq.read_table('images_hf/Detection/COCO/images_0001.parquet')
-df = table.to_pandas()
-# Create image lookup
-image_dict = {row['image_path']: row['image'] for _, row in df.iterrows()}
-# Load specific image
-img_path = "images_hf/Detection/COCO/000000178538.jpg"
-img_bytes = image_dict[img_path]
-image = Image.open(io.BytesIO(img_bytes))
-```
-### Loading Annotations
-```python
-import json
-with open('annotations_hf/Detection/coco_llava_train_epoch1_part1.jsonl', 'r') as f:
-    for line in f:
-        data = json.loads(line)
-        image_path = data['image']
-        conversations = data['conversations']
-        # Process annotation...
-```
-### Example: Training Pipeline
-```python
-import json
-import pyarrow.parquet as pq
-from PIL import Image
-import io
-# 1. Load parquet images
-table = pq.read_table('images_hf/Detection/COCO/images_0001.parquet')
-image_lookup = {row['image_path']: row['image']
-                for _, row in table.to_pandas().iterrows()}
-# 2. Read annotations
-with open('annotations_hf/Detection/coco_llava_train_epoch1_part1.jsonl') as f:
-    for line in f:
-        data = json.loads(line)
-        # Get image
-        img_bytes = image_lookup[data['image']]
-        image = Image.open(io.BytesIO(img_bytes))
-        # Parse conversations
-        for conv in data['conversations']:
-            if conv['from'] == 'gpt':
-                # Extract boxes: <ref>text</ref><box>x1 y1 x2 y2</box>
-                # Train your model...
-                pass
-```
-## Categories
-### Detection
-Object detection datasets with bounding boxes identifying object categories.
-### Grounding
-Phrase grounding datasets connecting natural language descriptions to image regions.
-### OCR
-Text detection and recognition in images.
-### UI
-User interface understanding including desktop, mobile, and web screenshots.
-### Layout
-Document layout analysis for forms, tables, and structured documents.
-### Pointing
-Point-based localization tasks.
-## Deduplication
-This dataset uses hash-based deduplication across all categories. Each unique image (by content) is stored only once, even if used in multiple datasets or categories. This reduces storage by approximately 30-50% compared to naive duplication.
-## Citation
-If you use this dataset in your research, please cite:
-```bibtex
-@dataset{locany_multimodal_2026,
-  title={Locany Multimodal Dataset},
-  author={Your Name},
-  year={2026},
-  publisher={Hugging Face},
-  url={https://huggingface.co/datasets/yk3701208/locany-dataset}
-}
-```
-## License
-This dataset is released under CC-BY-4.0 license.
-## Acknowledgments
-This dataset aggregates and processes data from multiple public sources including COCO, Object365, OpenImages, and others. Please see individual dataset licenses for specific restrictions.