Incremental Learning of NCM Forests for Large-Scale Image Classification

Abstract

In recent years, large image data sets such as "Ima-geNet", "TinyImages" or ever-growing social networks like "Flickr" have emerged, posing new challenges to image classification that were not apparent in smaller image sets. In particular, the efficient handling of dynamically growing data sets, where not only the amount of training images, but also the number of classes increases over time, is a relatively unexplored problem. To remedy this, we introduce Nearest Class Mean Forests (NCMF), a variant of Random Forests where the decision nodes are based on nearest class mean (NCM) classification. NCMFs not only outperform conventional random forests, but are also well suited for integrating new classes. To this end, we propose and compare several approaches to incorporate data from new classes, so as to seamlessly extend the previously trained forest instead of re-training them from scratch. In our experiments, we show that NCMFs trained on small data sets with 10 classes can be extended to large data sets with 1000 classes without significant loss of accuracy compared to training from scratch on the full data.

Iterative 3D shape classification by online metric learning

“…Thirdly, the scalable collections of 3D shape must be processed dynamically. The incremental learning and updating strategies have been widely used in the traditional classifier research, such as image classification (Ristin et al, 2014), object tracking (Collins et al, 2005) and shape segmentation . However, the way of classifying the collection of 3D shapes brings three aspects of new dynamic requirements: the size of the shape set might be too big to be handled at one time, the number of the remaining shapes after several recurrent grouping is too small to be categorized in the current context until enough shapes are added, and the new 3D shapes may be added dynamically to the collection.…”

Indexing Ensembles of Exemplar-SVMs with rejecting taxonomies

“…Liu et al [17] instead have proposed a probabilistic approach for learning the label tree parameters using maximum likelihood estimation. Similarly, random forests have been used to build fast hierarchies for classification [23] and fine-grained categorization [31].…”

IKNN-SVM: A Hybrid Incremental Algorithm for Image Classification

“…A set of experiments based on the open dataset Caltech256 [9] shows that (1) IKNN-SVM performs a higher recognition accuracy than traditional KNN and close to SVM with linear kernel. (2) In the scenarios of high real-time requirements, IKNN-SVM reacts more quickly compared with some other incremental methods.…”

“…Since the training data are usually open-ended and dynamic in real life, Incremental learning becomes increasingly important, especially in the practice of machine learning with regard to image classification [2]. For example, an image recognition cloud service should be able to utilize the ever-increasing data from Internet to improve its accuracy during its whole lifetime, and a robot should be able to learn from the human guidance as well as its own experience and incrementally extend its knowledge on object's images when it is performing a task.…”

“…Besides, KNN-based incremental learning algorithms has the following advantages as well: (1) Requiring little training time. (2) Effortless to deal with large and incremental data. (3) Avoiding overfitting of parameters, which is a central issue in learning based approaches.…”