Download files
Abstract
Object detection in images and videos is an important topic in computer vision. In general, a large number of training samples are required to train an object detector with decent accuracy. However, annotating training samples can be inflexible and expensive. Weakly supervised object detection (WSOD) can solve the problem by reducing the amount of manual annotation effort in training images. In this thesis, we propose our own WSOD approach based on multiple instance learning (MIL) and Boosting techniques and the experiments on real world images and video sequences demonstrate the advantages of the approach and its superior performances over the state-of-the-art approaches.
There are three contributions in this thesis. Firstly, we introduce a novel offline object detection approach which does not require manual annotation of training samples yet still has comparable discriminative power to supervised learning approaches. In the proposed approach, object hypotheses are annotated from images by a classical approach; then each object hypothesis' probability of being a true object of interest (denoted as soft label), is estimated as a multiple instance learning (MIL) problem; next, the object hypotheses and their soft labels are used to train an object detector based on a proposed Boosting algorithm.
Secondly, to take advantage of both weakly supervised learning and online learning, an online weakly supervised object detection approach is proposed. The soft labels of streaming data are estimated and then a proposed online Boosting algorithm is applied to construct and update a Boosting classifier with the streaming data and their soft labels.
Moreover, identifying positive instances in positive bags under MIL settings can be beneficial to WSOD. It can work as an automatic annotator and save a huge amount of manual annotation cost. It also provides a flexible way to train different classifiers based on the annotated instances for WSOD. To solve the MIL problem of labelling all instances in positive bags instead of just the bags alone, we propose a novel soft label estimation algorithm based on expectation maximization (EM).