---

PP-YOLOE

PP-YOLOE is an industrial-grade, state-of-the-art object detector designed for high performance and easy deployment. Built on the foundation of PP-YOLOv2, PP-YOLOE introduces several advanced features, including an anchor-free paradigm, a powerful backbone and neck with CSPRepResStage, ET-Head, and the dynamic label assignment algorithm TAL (Task Alignment Learning).

---

Key Features

• Anchor-Free Paradigm: Simplifies the detection pipeline and improves performance.
• CSPRepResStage: A powerful backbone and neck architecture for enhanced feature extraction.
• ET-Head: Efficient and accurate detection head.
• TAL (Task Alignment Learning): Dynamic label assignment algorithm for better training.
• High Performance: Achieves state-of-the-art results on benchmark datasets.
• Easy Deployment: Optimized for industrial applications with minimal setup.

---

Getting Started

1. Installation

To get started with PP-YOLOE, follow these steps:

1. Clone the repository:

   git clone https://github.com/Gaurav14cs17/YOLOE.git
   cd YOLOE
   

2. Install the required dependencies:

   pip install -r requirements.txt
   

3. Download the pretrained weights and configuration files from the PaddlePaddle Model Zoo.

2. Inference

Run inference on an image, video, or webcam using the provided scripts:

# Run on image
python inference.py \
    --source /path/to/image.jpg \
    --weights /path/to/model.pt \
    --model-size s \
    --conf-thres 0.25

# Run on video
python inference.py \
    --source /path/to/video.mp4 \
    --weights /path/to/model.pt

# Run on webcam (use webcam index)
python inference.py \
    --source 0 \
    --weights /path/to/model.pt

# Run on directory of images
python inference.py \
    --source /path/to/images/ \
    --weights /path/to/model.pt \
    --save-txt

3. Training

To train PP-YOLOE on your custom dataset:

COCO Format Dataset:

python train.py \
    --data-dir /path/to/coco \
    --train-ann /path/to/annotations/instances_train2017.json \
    --val-ann /path/to/annotations/instances_val2017.json \
    --model-size s \
    --epochs 300 \
    --batch-size 16 \
    --img-size 640 \
    --lr 0.01 \
    --amp  # Enable mixed precision training

YOLO Format Dataset:

python train.py \
    --data-dir /path/to/custom_data \
    --model-size s \
    --num-classes 10 \
    --epochs 100 \
    --batch-size 16

Training Options:

• --model-size: Model size (s, m, l, x)
• --epochs: Number of training epochs
• --batch-size: Batch size per GPU
• --img-size: Input image size
• --lr: Initial learning rate
• --amp: Enable automatic mixed precision
• --resume: Resume from checkpoint

---

Model Architecture

Backbone: CSPRepResStage

The backbone of PP-YOLOE is based on CSPRepResStage, which enhances feature extraction capabilities while maintaining efficiency.

Neck: ET-Head

The ET-Head is a lightweight and efficient detection head that improves accuracy and reduces computational overhead.

Anchor-Free Paradigm

PP-YOLOE adopts an anchor-free approach, which simplifies the detection pipeline and eliminates the need for anchor box tuning.

TAL (Task Alignment Learning)

The Task Alignment Learning (TAL) algorithm dynamically assigns labels during training, improving the alignment between classification and localization tasks.

---

Performance

PP-YOLOE achieves state-of-the-art performance on benchmark datasets like COCO. Below are some key metrics:

| Model | mAP (COCO) | FPS | |-----------------|------------|------| | PP-YOLOE-L | 50.9 | 78 | | PP-YOLOE-X | 52.2 | 68 |

---

Results

Object Detection Demo

Check out the object detection results using PP-YOLOE:

---

References

• PP-YOLOE Paper: PP-YOLOE: An Evolved Version of YOLO
• PaddlePaddle Detection Repository: GitHub Repository

---

Folder Structure

Here's an overview of the repository structure:

YOLOE/
├── models/                   # Model architecture
│   ├── __init__.py
│   ├── backbone.py           # CSPResNet backbone
│   ├── pan.py                # CustomCSPPAN neck (FPN + PAN)
│   ├── head.py               # PPYOLOEHead detection head
│   ├── yoloe.py              # Main YOLOE model
│   └── iou_loss.py           # IoU loss functions (GIoU, DIoU, CIoU)
├── assigners/                # Target assignment utilities
│   ├── __init__.py
│   ├── atss_assigner.py      # ATSS assigner
│   ├── bbox_utils.py         # Bounding box utilities
│   └── utils.py              # Helper functions
├── images/                   # Demo images and visualizations
│   └── yoloe.png
├── train.py                  # Training script
├── inference.py              # Inference script (image/video/webcam)
├── dataset.py                # Dataset classes (COCO, YOLO format)
├── requirements.txt          # Python dependencies
└── README.md                 # Project README file

Model Sizes

| Model | Width Mult | Depth Mult | Parameters | |-------|------------|------------|------------| | YOLOE-S | 0.50 | 0.33 | ~7M | | YOLOE-M | 0.75 | 0.67 | ~25M | | YOLOE-L | 1.00 | 1.00 | ~50M | | YOLOE-X | 1.25 | 1.33 | ~100M |

Dataset Formats

COCO Format

Standard COCO JSON annotation format with images, annotations, and categories.

YOLO Format

data/
├── train/
│   ├── images/
│   │   ├── img001.jpg
│   │   └── img002.jpg
│   └── labels/
│       ├── img001.txt
│       └── img002.txt
└── val/
    ├── images/
    └── labels/

Label file format (normalized coordinates):

class_id x_center y_center width height

---

How to Contribute

We welcome contributions! If you'd like to contribute to this project, please follow these steps:

1. Fork the repository.
2. Create a new branch for your feature or bug fix.
3. Commit your changes.
4. Submit a pull request.

---

License

This project is licensed under the Apache License 2.0. See the LICENSE file for details.

---

Acknowledgments

• Thanks to the PaddlePaddle team for developing and open-sourcing PP-YOLOE.
• Special thanks to the contributors of the PaddleDetection repository for their support and tools.

---