Exploring Convolutional Neuronal Networks for RGB-based Object Recognition

This project consists in developing a code able to recognize objects from images, the code can be then implemented on a drone or other system to enable a self-driving mode. In order to achieve this result the structure is based on a Convolutional Recursive Neural Network, CRNN, obtained from the fusion of two standard neural networks: Convolutional Neural Network (CNN) and Recursive Neural Network (RNN). The CNN is used as features extractor and its output is passed to the RNN that elaborate them exploiting its tree-like structure. The reason to join those two networks is that the RNN improves considerably the accuracy performance respect to a CNN alone. The dataset used in this work is the Washington RGBD which contains almost 8500 different images divided in 51 different classes representing objects. The CNN is a pretrained network and the RNN does not require any training since its weights are randomly initialized, the only part to be trained it is the softmax layer which is the final block with the role to select the final prediction. One of the critical steps it is the choice of the size of the RNN, because from this choice depends the performance in terms of accuracy and speed, the solution found at the end of the work has produced satisfactory results.