swimming-pool-detection's Introduction

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 (¬‿¬)).

Class Activation Maps on tiles
Input image
Heatmap generated by the CNN
Binarization and contour detection
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

Recommend Projects