Indoor-Outdoor Image Classification

Binary image classifier (indoor vs. outdoor scenes) trained on a curated subset of YouTube-8M frames using transfer learning on ResNet-50 with TensorFlow.

Approach

• Dataset: 169 curated images from YouTube-8M covering 12 scene categories (Bedroom, Bathroom, Classroom, Office, Living Room, Dining Room for indoor; Landscape, Skyscraper, Mountain, Beach, Ocean for outdoor)
• Model: ResNet-50 backbone (ImageNet pretrained) with fine-tuned classification head
• Techniques: Image augmentation, early stopping, efficient training on limited data
• Testing: Pytest suite covering data loading and augmentation pipelines

Project Structure

indoor_outdoor/
├── model.py                    # ResNet-50 model definition
├── training.py                 # Training loop
├── single_image_predictions.py # Inference on individual images
└── utils/
    ├── config.py
    ├── dataset.py
    ├── image_functions.py
    └── utils.py
data/                           # Image data (not committed)
references/                     # indoor_outdoor_images.zip (169 curated images)
tests/                          # Pytest suite

Quickstart

Requires Python 3.9+ and uv.

git clone https://github.com/bchadburn/computer-vision.git
cd computer-vision
uv venv && source .venv/bin/activate
uv pip install -r requirements.txt

Unzip the curated image set:

unzip references/indoor_outdoor_images.zip -d data/

Train the classifier:

python indoor_outdoor/training.py

Run inference on a single image:

python indoor_outdoor/single_image_predictions.py --image path/to/image.jpg

Run Tests

pytest tests/ -v