Keywords to visual categories: Multiple-instance learning forweakly supervised object categorization

By Sudheendra Vijayanarasimhan, Kristen Grauman. 2008

150 citations recorded by the source corpus.

Conventional supervised methods for image categorization rely on manually annotated (labeled) examples to learn good object models, which means their generality and scalability depends heavily on the amount of human effort available to help train them. We propose an unsupervised approach to construct discriminative models for categories specified simply by their names. We show that multiple-instance learning enables the recovery of robust category models from images returned by keyword-based search engines. By incorporating constraints that reflect the expected sparsity of true positive examples into a large-margin objective function, our approach remains accurate even when the available text annotations are imperfect and ambiguous. In addition, we show how to iteratively improve the learned classifier by automatically refining the representation of the ambiguously labeled examples. We demonstrate our method with benchmark datasets, and show that it performs well relative to both state-of-the-art unsupervised approaches and traditional fully supervised techniques.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Webly Supervised Learning of Convolutional Networks (2015, background): “We believe a more juicier direction is to feed more data into these networks (in fact, deeper networks would often need more data to train).”
Learning Object Categories From Internet Image Searches (2010, background): “Vijayanarasimhan and Grauman [ 42 ] use multiple-instance learning techniques to overcome the labeling noise in Internet images.”
Web-Based Classifiers for Human Action Recognition (2012, methodology): “Content-based image retrieval research has focused primarily on the retrieval of certain (and mostly unarticulated) objects; see [13]–[16] for some recent work.”
Object, Scene and Actions: Combining Multiple Features for Human Action Recognition (2010, unknown): “Multiple Instance Learning (MIL) paradigm has been explore d in quite a number of studies, both in machine learning( [2]) and computer visi on( [23, 31, 5]).”
Multi-Instance Learning With Emerging Novel Class (2021, methodology): “Additionally, in computer vision, [11] used MIL for multi-class image categorization, and [12] employed MIL to handle human part detection problems.”
Deep multiple instance learning for image classification and auto-annotation (2015, background): “Examples include object and face detection [43], visual categorization [42], segmentation [45] and image retrieval [24].”
Robust Object Tracking with Online Multiple Instance Learning (2011, unknown): “This technique has found recent success in other computer vision areas, such as object detection [14], [15] and object recognition [16], [17], [18].”
Finding Actors and Actions in Movies (2013, background): “Since we have partial knowledge about class membership of samples in bags, our problem is related to multiple instance learning (MIL) [24].”
WebVision Database: Visual Learning and Understanding from Web Data (2017, methodology): “Vijayanarasimhan and Grauman [38] proposed a multi-instance learning (MIL) method to explicitly handle the label noise when using web images as training data.”
Large-Scale Live Active Learning: Training Object Detectors with Crawled Data and Crowds (2011, background): “, (24; 22; 25)); however, such methods assume a single prominent object of interest per image, whereas for detection we will have cluttered candidate images that require a bounding box to identify the object.”
Learning the Relative Importance of Objects from Tagged Images for Retrieval and Cross-Modal Search (2012, methodology): “…al. 2009; Hwang and Grauman 2010b), using keyword search for inexpensive dataset creation (Fergus et al. 2005; Li et al. 2007; Schroff et al. 2007; Vijayanarasimhan and Grauman 2008), and building richer image representations based on the two simultaneous “views” for retrieval or clustering…”
Audio event and scene recognition: A unified approach using strongly and weakly labeled data (2016, background): “For example, multiple instance learning based weakly supervised image retrieval and visual object detection have been explored in computer vision community [19], [20], [21].”
Linking People in Videos with "Their" Names Using Coreference Resolution (2014, background): “All these works use the names from scripts as weak labels in a multiple instance learning (MIL) setting [40].”
Learning to recognize objects in egocentric activities (2011, methodology): “More recently, Multiple Instance Learning (MIL) has shown great promise as a method for weakly supervised learning in computer vision communities [1, 9, 5, 23, 6].”