Hypercolumns, UNet and GSUnet comparison
Comparison of multiple method to perform segmentation on Vaihingen dataset. Three models with increasing complexity were tested: Hypercolumns, UNet ans Gated-Shaped UNet.
Voice and Face matching using MobileNet
Classification of persons based on their voice and faces. To do so, multiple processes were compared from basic VGG network to transformer based MobileNet V2.
Flood Forecasting
Short study of multiple RNNs from the simple LSTM or GRU based model to more complex encoder-decoder with attention layers. In addition, an analysis of diverse data inputs, simple water levels up to complete weather and precipitations, are investigated to forecast the water level of one precise station.