Update README.md

README.md

@@ -43,3 +43,54 @@ configs:
   - split: test
     path: data/test-*
 ---
+
+# Brazilian Jiu-Jitsu (BJJ) Positions Dataset
+
+## Overview
+This dataset contains annotated images for **Brazilian Jiu-Jitsu (BJJ) positions** such as guard, mount, side control, turtle, takedown, and standing.  
+Each annotation provides **category labels**, **bounding boxes**, and **keypoints** in COCO-style format.  
+
+The dataset has been pre-split into:
+- **train** (80%)
+- **validation** (10%)
+- **test** (10%)
+
+Splits were performed **per-category by percentage** (so each class is represented proportionally in all splits).
+
+---
+
+## Data Format
+Each sample contains the following fields:
+
+- `image_path`: Local path to the image (may not be valid on Hub, provided for reference).  
+- `file_name`: File name of the image.  
+- `image_id`: Unique image identifier.  
+- `annotation_id`: Unique annotation identifier.  
+- `category_id`: Integer ID for the BJJ position.  
+- `category_name`: Human-readable label (e.g. `"mount1"`, `"closed_guard2"`, `"standing"`).  
+- `keypoints`: 17 keypoints `[x, y, v]` (COCO style; `v=0` not labeled, `1` labeled but not visible, `2` visible).  
+- `bbox`: `[x, y, w, h]` bounding box.  
+- `iscrowd`: Standard COCO flag.  
+- `area`: Area of the bounding box.  
+- `image_url` *(optional)*: If provided, a URL for downloading the image.
+
+---
+
+## Statistics
+Approximate distribution (train/val/test combined):
+
+- Total samples: ~196k  
+- Number of classes: 18 (BJJ positions)  
+- Example classes:  
+  - `standing`  
+  - `mount1`, `mount2`  
+  - `closed_guard1`, `closed_guard2`  
+  - `turtle1`, `turtle2`  
+  - `5050_guard`  
+  - `side_control1`, `side_control2`  
+  - `open_guard1`, `open_guard2`  
+  - `takedown1`, `takedown2`  
+  - `half_guard1`, `half_guard2`  
+  - `back1`, `back2`
+
+---