Abstract

The aim of this PhD thesis is to make a step forward towards teaching computers to understand videos in a similar way as humans do. In this work we tackle the video classification and/or action recognition tasks. This thesis was completed in a period of transition, the research community moving from traditional approaches (such as hand-crafted descriptor extraction) to deep learning. Therefore, this thesis captures this transition period, however, unlike image classification, where the state-of-the-art results are dominated by deep learning approaches, for video classification the deep learning approaches are not so dominant. As a matter of fact, most of the current state-of-the-art results in video classification are based on a hybrid approach where the hand-crafted descriptors are combined with deep features to obtain the best performance. This is due to several factors, such as the fact that video is a more complex data as compared to an image, therefore, more difficult to model and also that the video datasets are not large enough to train deep models with effective results. The pipeline for video classification can be broken down into three main steps: feature extraction, encoding and classification. While for the classification part, the existing techniques are more mature, for feature extraction and encoding there is still a significant room for improvement. In addition to these main steps, the framework contains some pre/post processing techniques, such as feature dimensionality reduction, feature decorrelation (for instance using Principal Component Analysis - PCA) and normalization, which can influence considerably the performance of the pipeline. One of the bottlenecks of the video classification pipeline is represented by the feature extraction step, where most of the approaches are extremely computationally demanding, what makes them not suitable for real-time applications. In this thesis, we tackle this issue, propose different speed-ups to improve the computational cost and introduce a new descriptor that can capture motion information from a video without the need of computing optical flow (which is very expensive to compute). Another important component for video classification is represented by the feature encoding step, which builds the final video representation that serves as input to a classifier. During the PhD, we proposed several improvements over the standard approaches for feature encoding. We also propose a new feature encoding approach for deep feature encoding. To summarize, the main contributions of this thesis are as follows3: (1) We propose several speed-ups for descriptor extraction, providing a version for the standard video descriptors that can run in real-time. We also investigate the trade-off between accuracy and computational efficiency; 
(2) We provide a new descriptor for extracting information from a video, which is very efficient to compute, being able to extract motion information without the need of extracting the optical flow; (3) We investigate different improvements over the standard encoding approaches for boosting the performance of the video classification pipeline.;(4) We propose a new feature encoding approach specifically designed for encoding local deep features, providing a more robust video representation.