A course project using vgg for swimming pool detection.
A jupyter notebook implementing image preprocessing, vgg training and evaluation. Since there is no access to GPUs, I made use of three memory control approaches:
- train_test_split on file names and load images when feeding into the model;
- resize into smaller images before training;
- apply "imagenet" as the initial weights.
Here below shows the best and the worst detections with prediction in red and true box in blue.
Examples of the input satellite images, the original folder containing 14964 images.
A dictionary of the coordinates of swimming pool for each image.