Automatic annotation for weakly supervised learning of detectors

Abstract

Object detection in images and action detection in videos are among the most widely studied computer vision problems, with applications in consumer photography, surveillance, and automatic media tagging. Typically, these standard detectors are fully supervised, that is they require a large body of training data where the locations of the objects/actions in images/videos have been manually annotated. With the emergence of digital media, and the rise of high-speed internet, raw images and video are available for little to no cost. However, the manual annotation of object and action locations remains tedious, slow, and expensive. As a result there has been a great interest in training detectors with weak supervision where only the presence or absence of object/action in image/video is needed, not the location. This thesis presents approaches for weakly supervised learning of object/action detectors with a focus on automatically annotating object and action locations in images/videos using only binary weak labels indicating the presence or absence of object/action in images/videos. First, a framework for weakly supervised learning of object detectors in images is presented. In the proposed approach, a variation of multiple instance learning (MIL) technique for automatically annotating object locations in weakly labelled data is presented which, unlike existing approaches, uses inter-class and intra-class cue fusion to obtain the initial annotation. The initial annotation is then used to start an iterative process in which standard object detectors are used to refine the location annotation. Finally, to ensure that the iterative training of detectors do not drift from the object of interest, a scheme for detecting model drift is also presented. Furthermore, unlike most other methods, our weakly supervised approach is evaluated on data without manual pose (object orientation) annotation. Second, an analysis of the initial annotation of objects, using inter-class and intra-class cues, is carried out. From the analysis, a new method based on negative mining (NegMine) is presented for the initial annotation of both object and action data. The NegMine based approach is a much simpler formulation using only inter-class measure and requires no complex combinatorial optimisation but can still meet or outperform existing approaches including the previously pre3 sented inter-intra class cue fusion approach. Furthermore, NegMine can be fused with existing approaches to boost their performance. Finally, the thesis will take a step back and look at the use of generic object detectors as prior knowledge in weakly supervised learning of object detectors. These generic object detectors are typically based on sampling saliency maps that indicate if a pixel belongs to the background or foreground. A new approach to generating saliency maps is presented that, unlike existing approaches, looks beyond the current image of interest and into images similar to the current image. We show that our generic object proposal method can be used by itself to annotate the weakly labelled object data with surprisingly high accuracy.