ECCV 2018 Daily - Monday

Learning to Detect and Track Visible and Occluded Body Joints in a Virtual World Matteo Fabbri is a PhD student from the University of Modena and Reggio Emilia in Italy. His supervisor is Rita Cucchiara . He speaks to us ahead of his poster session today. Matteo tells us that the work he is presenting is about multiple people tracking using pose estimation . It proposes using a computer graphic video game like Grand Theft Auto to collect a huge dataset and train the annotation. It is a new technique and one they exploit in order to perform tracking in a shorter span. Matteo explains further: “ In practical terms, it is a deep neural network that is extended by a well-known network by Cao et al. from CVPR 2017. There is a feature extractor, VGG-19, that extracts features. After that we output four branches. One branch that detects the joints; one that detects the limbs that connects those joints onto people; one for occluded heatmaps, so we are able to detect occluded body parts of people; and the last branch that connects joints of the same person across time. With this branch we are able to perform tracking .” Matteo hopes that the dataset will be downloaded and used, because of the vast amount of trained and occlusion information it contains. He adds that the next step will be to go from 2D to 3D . To find out more about Matteo’s exciting work, visit his poster today at 16:00 [P-1B-07]. “It is a deep neural network that is extended by a well- known network by Cao et al. from CVPR 2017.” Daily MONDAY 10 Matteo Fabbri 15