The challenge of HAhRD project is to implement new algorithms to classify objects from 3D images-like coming from the data acquisition of the future sub-detector of CMS. This detector which contains about 6 million channels will be used to reconstruct the 3D cluster-like objects coming from hundreds of impinging particles arising from the proton-proton collisions within the Large Hadron Collider. Previous successful studies using Convolution Neural Networks (CNN), in particular in HSF-GSOC program context, have been done for single object. In this proposal we want to focus on object detection models to scan all clusters in our 3D image-like. Recent cutting-edge models like R-CNN, fast(er) R-CNN, YOLO, etc. are in the top of the ranking of the ImageNet competition. One of these models, called Mask R-CNN model have already been evaluated in HAhRD project with 2D projections (of our 3D data), thanks to a published implementation. We propose for GSOC’19 to derivate a 3D version (true 3D input “gray” images), from the original 2D (RGB) Mask R-CNN implementation.
Good knowledge on object detection (in particular R-CNN) and Deep Learning, very good skills in TensorFlow and Keras, good python skills. Basic Knowledge in physics would be appreciated.