Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.784240
Title: Video-based object recognition
Author: Liu, Yang
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2017
Availability of Full Text:
Access from EThOS:
Access from Institution:
Abstract:
With the number of videos growing rapidly in modern society, automatically recognizing objects from video input becomes increasingly pressing. Videos contain abundant yet noisy information, with easily obtained video-level labels. This thesis focuses on the task of video-based object recognition. The key problem it tackles is how to effectively accumulate the useful information in a video for object recognition, especially when only the effortless video-level annotation is available. This thesis follows the classic two-step computer vision algorithm pipeline of building object representation and learning classification model. First, two novel set-of-sets representations are proposed to represent a video sequence. The respective matching schemes are also designed to make use of the proposed representations for video matching. Then, a novel algorithm, latent bi-constraint SVM (LBSVM), is proposed to learn the classification model for video-based object recognition. It consists of two new constraints and a latent variable for SVM to effectively accumulate the abundant information in a video, by selecting the useful information and filter out the noise. Finally, a weakly supervised learning algorithm is proposed for video-based object recognition, simply using the weakly annotated video-level label. The location of the target object is inferred during the training iterations, which also helps to increase the object recognition accuracy. Since there are no suitable video datasets publicly available for video-based object recognition, four video datasets are collected in the thesis to evaluate the proposed methods: Office-Instance video dataset, Office-Category video dataset, Museum video dataset and Youtube video dataset. The four video datasets not only cover different research and application scenarios, but also illustrate the evolution of my research on the problem of video-based object recognition. The experimental results on the four video datasets demonstrate the benefits of the proposed methods.
Supervisor: Kim, Tae-Kyun Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.784240  DOI:
Share: