README.md

---
task_categories:
- object-detection
pretty_name: COCO 2014 DensePose Relabeling with Body Parts
---

# COCO 2014 DensePose Relabeling with Body Parts

This dataset is formatted for Ultralytics YOLO and is ready for training.
IMPORTANT !!!! Update the paths in the yaml inside the dataset folder 





## Demo

Here is what inference looks like: 
![image/jpeg](https://cdn-uploads.huggingface.co/production/uploads/66f45a63ea609a69012a1b39/vlhDa9SOzfoFmAo-Hc8dA.jpeg)


## Based on:
- [GitHub Repository](https://github.com/FraunhoferIKS/smf-object-detection?tab=License-1-ov-file#readme)
- [Paper](https://arxiv.org/pdf/2307.04533)

## Classes:
```json
{
    1: "Person",
    2: "Torso",
    3: "Hand",
    4: "Foot",
    5: "Upper Leg",
    6: "Lower Leg",
    7: "Upper Arm",
    8: "Lower Arm",
    9: "Head"
}
```

## Dataset Structure:
```
dataset/
 ├── images/
 │    ├── train/         # Training images (e.g., .jpg or .png)
 │    └── val/           # Validation images
 ├── labels/
 │    ├── train/         # Training labels (one .txt per image)
 │    └── val/           # Validation labels
 └── data.yaml           # YAML config file
```
In images/train threre are 13483 images which contain people, and also 1000 images which contain no people which are called backgrounds (they help avoid False Positives).
In images/val there are 2215 images which contain people

In labels/train there are 13482 txt files which contain object detection information in the ultralytics yolo format. The 1000 background images have txt files (they don't need it)
In labels/val there are 2215 txt files


## Training Example:
```python
from ultralytics import YOLO
import os

os.environ["CUDA_VISIBLE_DEVICES"] = "0"  

model = YOLO("yolo11l.pt")  

results = model.train(data="dataset/data.yaml", epochs=50, imgsz=640, device="0", batch=16, save=True)
```