Authors:
[Edition 2024] : Kamil Hanna, Enrico Tirri,Giorgio Negro,
[Edition 2023] : Kamil Hanna, Filippo Desantis, Livia Giacomin
Note that additional information about the models are present in the report pdf files ,in each directory. Moreover, please note that some of the competition links might be down/unaccessible to users without Politecnico di Milano credentials.
Competition available on Codabench.
This challenge focuses on building a Convolutional Neural Network (CNN) classifier to categorize 96x96 RGB images of blood cells into one of eight predefined classes, each representing a specific cell state. The cleaned dataset comprises 11,959 labeled images after removing significant outliers and handling class imbalance.
The final solution is based on ConvNeXtBase, pre-trained on ImageNet, combined with a custom classification head including SELU-activated dense layers, dropout, and L1L2 regularization. We applied an extensive data augmentation pipeline, including MixUp, CutMix, RandAugment, and channel shuffling, with augmentations pre-applied to avoid performance issues during training.
All code and assets are located in the Challenge1_2024 folder, which also includes a detailed report outlining the full pipeline, augmentation techniques, architecture design, and tested alternatives. In addition to the final model, the folder contains our two best alternative models: one based on EfficientNetV2B3 and another using a simple custom convolutional network. The final ConvNeXt model achieved a Codabench score of 0.86, with performance validated across training, validation, and local test sets.
Competition available on Kaggle.
This challenge focuses on a semantic segmentation task, classifying 64x128 greyscale images of Mars terrain into five classes, each corresponding to a specific type of terrain. The classifier must assign a class label to every pixel in the input image.
The dataset contains 2,615 images, reduced to 2,505 after removing outliers detected via t-SNE analysis. To address class imbalance, especially for class 4, we applied targeted data augmentation, generating up to 36 variations per sample for that class. The training-validation split was carefully designed to reflect the imbalance, while optimizing data efficiency.
Our final model is based on U-Net 3+, modified with architectural enhancements including dropout layers, trainable skip connection filters, squeeze-and-excite blocks, and L2 regularization. We used a weighted sparse categorical loss function and trained the model over two 100-epoch sessions, reaching a mean IoU of 0.643 on the Kaggle leaderboard. All training used a batch size of 64, with AdamW optimizer and learning rate scheduling.
All code and resources are located in the Challenge2_2024 folder, including:
- A detailed report describing the preprocessing pipeline, architecture experiments (MS-UNet, MarsSeg, Deep Residual U-Net), and training results.
- Python scripts used for dataset modification and augmentation.
- A reference implementation of a basic U-Net model for comparison and reproducibility.
Competition available on Codalab.
This challenge addresses a binary classification task: identifying whether a leaf in an image is healthy or unhealthy. The dataset consists of 5200 RGB images of size 96x96, each labeled accordingly.
After cleaning the data—removing 98 duplicate outliers—and balancing the classes using SMOTE, we trained a deep learning model using transfer learning. The final architecture is based on ConvNeXt Large, with the first 90 layers frozen, and incorporates batch normalization, dropout (0.1), and a cyclical learning rate. The best model achieved a validation accuracy of 0.81.
All code is located in the Challenge1_2023 folder, along with a detailed report outlining the full pipeline and additional architecture experiments.
Competition available on Codalab.
This challenge addresses a time series forecasting task: predicting the next steps of given time series sequences. The dataset contains 48,000 multivariate time series, each with 2,776 time steps, and associated metadata specifying a start and end index for the valid range. Although category labels (A–F) were provided, they were not used, as they did not correspond to distinct series behaviors.
After extracting the valid segments from each series, we processed the data into fixed-length sequences of 200 steps, paired with targets of 18 future steps, using a custom splicing approach. We applied basic data augmentation, including duplication with slight rounding variations, which improved results.
The final model is a simple LSTM-based neural network without convolutional or attention layers, followed by three dense layers. It was trained using a batch size of 32 and MSE loss, and evaluated using validation MSE. The best submission achieved an MSE of 0.009 on the CodaLab leaderboard.
All code for this project is located in the Challenge2_2023 folder, along with a detailed report documenting the full pipeline, architecture selection, and experimental results.