Giovani in un'ora - Ciclo di seminari - Seconda parte
- Day - Time: 21 November 2018, h.11:00
- Place: Area della Ricerca CNR di Pisa - Room: C-29
Fabio Carrara - Real-Time Action Detection and Prediction in Human Motion Streams
Abstract: Motion capture data digitally represent human movements by sequences of 3D skeleton configurations. Such spatiotemporal data, often recorded in the stream-based nature, need to be efficiently processed to detect actions of interest, such as real-time hand gestures for human-computer interaction. Alternatively, parts of a continuous stream can be persistently stored and then automatically annotated to become searchable, and thus reusable for future retrieval or pattern mining. In this talk, we focus on multi-label segmentation and annotation of user-specified actions in sequences as well as continuous streams of motion capture data. In particular, we show how the adoption of recent advances in recurrent neural networks (RNNs) effectively encodes the skeleton frames in the hidden network states and enable effective action segmentation. We extensively evaluate our models on the three use cases of real-time stream annotation, offline annotation of long sequences, and early action detection and prediction. Experiments show that leveraging RNNs we are able to obtain state-of-the-art performance in terms of effectiveness while being one order of magnitude more efficient, having the ability to annotate up to 10k motion frames per second.
Ioanna Miliou - Supermarket retail data as a proxy for predicting seasonal influenza
Abstract: In this work we use supermarket retail data as a proxy for predicting seasonal influenza. When the flu season arrives, people are starting to get sick. Getting sick affects their everyday life and behaviour. This change in behaviour propagates in their purchases in the supermarket. They will buy products that will reflect the fact that they are sick. We collect this information studying the purchases of each week in order to detect the more correlated products with the influenza adoption trend. We identify the customers that buy them during the influenza peak, and through them we identify the sentinels, frequent baskets of such customers during the peak. We use the weekly values, time series, of the next season for these sentinels as a proxy for the next flu season. Finally, using a regression model we assume that this week’s influenza depends on the influenza in previous weeks as well as the proxy in the current week and previous weeks. Our approach produces nowcasting and forecasting values for up to 4 weeks ahead.
Leonardo Robol - Markov chains and matrix functions
Abstract: Several measures of interest in the context of reliability assessment and performance measurement using Markov chains can be naturally expressed through the framework of matrix functions and tensors. This opens the door to a broad class of approximation methods, which can be used to improve state of the art techniques in the field. We discuss how to draw this connection, and how to exploit it numerically.