Glove Detection Project

This project uses YOLOv8 to detect hands with and without gloves in images. It can process a directory of images and will output annotated images as well as JSON files with detection details.

Demo

AsadNizami/glove-detection

Sample Output

Input Image	Detected Output

Installation

This project uses uv for package management.

1. Install uv:

   pip install uv

Usage

To run glove detection on a folder of images, use the main.py script.

uv run main.py <path_to_your_image_folder>

For example, to process the sample images provided:

uv run main.py sample_images

The annotated images will be saved in the output/ directory, and the detection logs in JSON format will be saved in the logs/ directory.

Dataset

The dataset used for this project is a custom dataset focused on glove detection, originally sourced from Roboflow: Hang and Glove Detect. It contains images of hands, which will be categorized into 'gloved_hand' and 'bare_hand' after cleaning the dataset.

Model

The primary model used for detection is YOLOv8s (YOLOv8 small), chosen for its balance of performance and efficiency. This model was fine-tuned from a pretrained yolov8s.pt checkpoint.

Preprocessing and Training

The training process involved several steps:

1. Data Cleaning: Initial data cleaning was performed using clean_dataset.py to remove duplicates or corrupted entries and to rename the classes.
2. Data Splitting: The dataset originally had train set and test set. Validation set was created from 20% of the train set.
3. Training: The YOLOv8s model was trained for 75 epochs using the train.py script. Key hyperparameters were configured as follows (from runs/detect/train3/args.yaml):
   • epochs: 75
   • batch: 16
   • imgsz: 640
   • model: yolov8s
   • cos_lr: True (cosine learning rate scheduler)
   • mixup: 0.1
   • flipud: 0.5 (vertical flip augmentation)
   • fliplr: 0.5 (horizontal flip augmentation)

What Worked and What Didn't

• YOLOv8s Performance: The YOLOv8s model demonstrated strong performance, achieving high Mean Average Precision (mAP) scores on the validation set.
• Hyperparameter Tuning: Extensive hyperparameter tuning was conducted across different training runs (train, train2, train3, train4).
• Model Comparison: While YOLOv8m (from train4) showed marginally higher mAP50-95, YOLOv8s (from train3) was selected as the preferred model due to its comparable performance and significantly smaller size, making it more efficient for deployment.
• Initial Models: Earlier training runs with YOLOv8n models showed lower performance compared to the 's' and 'm' variants, indicating the benefit of using larger models for this dataset.