Example Solutions - Konnsy/REAML2022-hackathon GitHub Wiki

Test data sets

Example for Task 1

If you want to see how Task 1 can be solved, have a look at the example code for Task 1. You can also load the trained model to use the given solution without training yourself.

The architecture

The network architecture is similar to a VGG-Net with a global maximum pooling instead of just flattening the results. This way, the model is independent of image sizes. The learning rate was adapted to match the implemented architecture. The other parts of the code frame, including the preprocessing, are unchanged.

Results

The given example solution reaches an accuracy of 94.5% on the validation data (based on per-image classification) and classifies each test set correctly (determined by a majority vote across all images of a set).

Example for Task 2

Example solution and trained model coming soon.

You can also access an example code and the corresponding trained model for Task 2. The contained segmenter is a ResNet-structure of 100 layers, each with 8 filters per convolution. The training employs random scalings of the image sizes and a random contrast adjustment.

The architecture

The detector is basically calling the blob detector from the open cv library. This part is a non-learning algorithm with few adjustable parameters fitting the detection of blobs in the segmented image.

Results

Having trained the model for 15 epochs, an accuracy of around 70% is achieved on validation and test data.

Visualizations of example detection (like those in the images below) can be found at https://tu-dortmund.sciebo.de/s/8VoNUFZQU5o1bHA.