CSCI2950P Final Project Report: Unsupervised Learning of Scene Categories
نویسنده
چکیده
We present the Hierarchical Latent Dirichlet Allocation (hLDA) for organizing a collection of scene images into a hierarchy in unsupervised fashion. The bag of visual words model with spatial pyramids is adpated to the hLDA where the spatial generality is used as a prior of assignments of topics to visual words. Both depth and width of the tree are unbounded and inferred by training data using the Nested Chinese Restaurant Process (nCRP). We use a collapsed Gibbs sampler as the posterior inference algorithm that use the Metropolis-Hastings (MH) steps to obtain next values of hyperparameters. We experiment our algorithm on both toy data set and scene dataset. The results show that the unsupervised learning method can also achieve a decent classification rate.
منابع مشابه
EE/CS148, Object Recognition, Final Project Report
There are clear distinctions between the object “car” and “cat”, but how are these labels assigned? It is easy for a person to group images into different labeled categories, but far less so for a computer. If a computer did possess the ability to classify unlabeled images (or parts of images), it could be applied to a number of technologies; including true image search engines and automatic me...
متن کاملCPSC 503 - Final project Part-of-speech filtering in unsupervised learning to improve discourse relation recognition
I evaluate a technique for improving the accuracy of discourse relation recognition by unsupervised classifiers that involves filtering the input features based upon their parts of speech. I report on experiments on various corpora and training set sizes in which classifiers trained on filtered features are less accurate than equivalent classifiers trained on unfiltered features.
متن کاملFinal Project: Parallel unsupervised learning with neural data
This note is an accompaniment to my poster for my class project. Despite its seeming length, it attempts to be as brief as is reasonable, explaining the figures in greater detail and providing additional computational details most relevant to the subject of the course.
متن کاملScene Classification Via pLSA
Given a set of images of scenes containing multiple object categories (e.g. grass, roads, buildings) our objective is to discover these objects in each image in an unsupervised manner, and to use this object distribution to perform scene classification. We achieve this discovery using probabilistic Latent Semantic Analysis (pLSA), a generative model from the statistical text literature, here ap...
متن کاملCategory co-occurrence modeling for large scale scene recognition
Scene recognition involves complex reasoning from low-level local features to high-level scene categories. The large semantic gap motivates that most methods model scenes resorting to mid-level representations (e.g. objects, topics). However, this implies an additional mid-level vocabulary and has implications in training and inference. In contrast, the semantic multinomial (SMN) represents pat...
متن کامل