• Object Detection
• Deep Learning
• Autonomous Grasping
• Robotics

The purpose of this thesis is to explore solutions to vision-based grasp estimation problem using Deep Learning techniques.
Fizyr 's implementation of the state-of-the-art Object Detector RetinaNet has been used, as well as the two open source Cornell and DeepGrasp datasets.
Starting from these, two datasets for training and evaluation has been derived: the Cornell-DeepGrasp and the Mixed-Mixed dataset. Both advantages and disadvantages of training process with different datasets have been discussed using Performance Curve, Evaluation Techniques and Bias Analysis.
At the end of this analysis the Mixed-Mixed dataset has been discarded due to an observable bias of the training set that compromises performance. Instead, training with Cornell dataset has shown not only good performance but also the possibility to achieve improvements.
The main contribution of this work is a structured approach to analyse performance of a detection network: in particular it is shown the usage of the RetinaNet architecture to perform a grasp proposal task and the estimation of information content of annotations and detections, using big data collection.
The algorithm has been implemented and tested on a UR5 Robot in order to perform a Bin-Picking task.