The amount of user generated video content is growing enormously with the increase in availability and affordability of technologies for video capturing (e.g. camcorders, mobile-phones), storing (e.g. magnetic and optical devices, online storage services), and sharing (e.g. broadband internet, social networks). It has become a common sight at social occasions like parties, concerts, weddings, vacations that many people are shooting videos at approximately the same time. Such concurrent recordings provide multiple views of the same event. In professional video production, the use of multiple cameras is very common. In order to compose an interesting video to watch, audio and video segments from different recordings are mixed into a single video stream. However, in case of non-professional recordings, mixing different camera recordings is not common as the process is considered very time consuming and requires expertise to do. In this thesis, we research on how to automatically combine multiple-camera recordings in a single video stream, called as a mashup. Since non-professional recordings, in general, are characterized by low signal quality and lack of artistic appeal, our objective is to use mashups to enrich the viewing experience of such recordings. In order to define a target application and collect requirements for a mashup, we conducted a study by involving experts on video editing and general camera users by means of interviews and focus groups. Based on the study results, we decided to work on the domain of concert video. We listed the requirements for concert video mashups such as image quality, diversity, and synchronization. According to the requirements, we proposed a solution approach for mashup generation and introduced a formal model consisting of pre-processing, mashupcomposition and post-processing steps. This thesis describes the pre-processing and mashup-composition steps, which result in the automatic generation of a mashup satisfying a set of the elicited requirements. At the pre-processing step, we synchronized multiple-camera recordings to be represented in a common time-line. We proposed and developed synchronization methods based on detecting and matching audio and video features extracted from the recorded content. We developed three realizations of the approach using different features: still-camera flashes in video, audio-fingerprints and audio-onsets. The realizations are independent of the frame rate of the recordings, the number of cameras and provide the synchronization offset accuracy at frame level. Based on their performance in a common data-set, audio-fingerprint and audio-onset were found as the most suitable to apply in generating mashups of concert videos. In the mashup-composition step, we proposed an optimization based solution to compose a mashup from the synchronized recordings. The solution is based on maximizing an objective function containing a number of parameters, which represent the requirements that influence the mashup quality. The function is subjected to a number of constraints, which represent the requirements that must be fulfilled in a mashup. Different audio-visual feature extraction and analysis techniques were employed to measure the degree of fulfillment of the requirements represented in the objective function. We developed an algorithm, first-fit, to compose a mashup satisfying the constraints and maximizing the objective function. Finally, to validate our solution approach, we evaluated the mashups generated by the first-fit algorithm with the ones generated by two other methods. In the first method, naive, a mashup was generated by satisfying only the requirements given as constraints and in the second method, manual, a mashup was created by a professional. In the objective evaluation, first-fit mashups scored higher than both the manual and naive mashups. To assess the end-user satisfaction, we also conducted a user study where we measured user preferences on the mashups generated by the three methods on different aspects of mashup quality. In all the aspects, the naive mashup scored significantly low, while the manual and first-fit mashups scored similarly. We can conclude that the perceived quality of a mashup generated by the naive method is lower than first-fit and manual while the perceived quality of the mashups generated by first-fit and manual methods are similar.
|Qualification||Doctor of Philosophy|
|Award date||19 Nov 2009|
|Place of Publication||Eindhoven|
|Publication status||Published - 2009|