Swimming Pool Detection on Aerial Images

One-day project ✍(◔◡◔)

Problematic

Detect pools on aerial images. Can be used for market analysis for a swimming pool seller or fight against tax evasion of undeclared swimming pools: 12,000 illegal swimming pools detected in Provence thanks to artificial intelligence.

Proposed Approach

Data

• 50x50px tiles containing a pool or not.
• ~1500 tiles for positive class, ~1500 for negative class.
• No segmentation masks, no coordinates. Just pool or no pool for each tile.
• Private dataset ¯\_(ツ)_/¯

Algorithm

• Convolutional Neural Network (resnet50) with class activation maps to generate a heatmap. Simple computer vision algorithms to detect "blobs" and return the swimming pools' coordinates.
• Inference is done on 800x1600px images. No need to split the image into 50x50 patches since the network uses a Global Average Pooling layer.
• Network is trained with simple data aug, Adam (with lr=3e-4 (¬‿¬)).

1. Class Activation Maps on tiles
2. Input image
3. Heatmap generated by the CNN
4. Binarization and contour detection
5. Draw bounding boxes

Results

Result 1

Result 2

Result 3

Result 4

Result 5

Some Remarks

• HSV filtering on blue shades before feeding a tile in the CNN might improve the results...?
• Binarization and contour detection need some finetuning.

Repo structure

├── README.md
├── pool_detection_example_notebook.ipynb     <- notebook for inference with some visualizations
├── models                                    <- model weights
├── src
│   ├── pool_detector.py                      <- inference
│   ├── data
│   │   └── split_dataset.py                  <- generate train/test/valid sets
│   ├── models
│   │   ├── predict_cnn.py                    <- check results
│   │   └── train_cnn.py                      <- train cnn