camera-occlusion 0.1.0

A library for applying realistic camera occlusion effects like rain and dust for computer vision data augmentation.

Realistic Camera Occlusion Effects for Computer Vision

This project provides a Python library and tools for applying realistic, parameterized camera occlusion effects like rain and dust to images. It's designed for data augmentation in computer vision tasks, such as training a traffic sign classifier to be more robust against adverse conditions.

Quick Start

pip install ...
obscure-image GTSRB_dataset/GTSRB/Final_Training/Images/00000/00000_00000.ppm  out.ppm --effect rain

Features

• Rain Effect: Simulates raindrops on a camera lens with distortion, shading, and highlights.
• Dust Effect: Simulates fine dust specks, scratches, and semi-transparent grime splotches.
• PyTorch Integration: Includes a Dataset class for on-the-fly data augmentation during model training.
• Command-Line Tools: Apply effects to single images or train a model directly from the command line.

Installation

1. Clone the repository:

   git clone [https://github.com/your-username/german_signs.git](https://github.com/your-username/german_signs.git)
   cd german_signs
   

2. Create and activate a virtual environment (recommended):

   python3 -m venv venv
   source venv/bin/activate
   

3. Install the required packages:

   pip install -r requirements.txt
   

Usage

There are three main ways to use this project:

1. As a Python Library

You can easily import and use the Rain and Dust effect classes in your own scripts.

import imageio.v2 as imageio
from camera_occlusion import Rain

# Load an image
image = imageio.imread("path/to/your/image.jpg")

# Apply a heavy rain effect
heavy_rain = Rain(num_drops=100, radius_range=(4, 8))
augmented_image = heavy_rain(image)

# Save or display the result
imageio.imwrite("rainy_image.jpg", augmented_image)

2. Gallery Demonstration

To see a gallery of all available effects and presets, run the example script. This will generate a plot showing various levels of rain and dust.

python examples/run_gallery.py

3. Training a Model

The train_model.py script trains a simple CNN on the German Traffic Sign Recognition Benchmark (GTSRB) dataset. You can specify different augmentation effects for the training and validation sets.

1. Download the data: You must first download the GTSRB dataset and place the training images in a GTSRB/ directory. You can find the data here: GTSRB Dataset. Your structure should look like GTSRB/Final_Training/Images/00000/....

2. Run training: The following command trains a model on images with "heavy dust" and evaluates it on images with "light rain".

   python train_model.py \
       --data-dir GTSRB/Final_Training/Images/ \
       --epochs 15 \
       --train-effect heavy_dust \
       --eval-effect light_rain \
       --save-path models/dusty_model.pth
   

License

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