Computer Vision News - December 2018

interest, so higher-order dependencies among pixels can be modeled to enforce the holistic segmentation of object instances. (b) The iterative algorithm performs video object segmentation efficiently. The algorithm alternates between a temporal step (between video frames) and a spatial step using a CNN along the image-space to refine segmentation results. 4. Background: The idea of applying CNNs in combination with MRF/CRF models is not new: DeepLab semantic segmentation framework attempted to improve the semantic labelling results produced by CNN, by using fully-connected CRF post-processing. The Jang and Kim video object segmentation method combined the outputs of a triple-branch CNN using MRF optimization. These loosely-combined algorithms did not, however, take full advantage of the MRF/CRF models’ strengths. Schwing and Urtasun jointly trained CNN and MRF by back-propagating gradient obtained during the MRF inference to CNN, however, no distinct improvement over separate training was achieved. The CRF-RNN model used a mean-field to approximate the CRF inference, within an RNN, arriving at an end-to-end trainable deep network, which greatly boosted performance. Deep Parsing Network (DPN) is an attempt to use a mean-field to approximate MRF inference in one training pass. The authors’ paper is trying to model higher-order potentials in MRFs with CNNs. Computer Vision News Research 5 Research Computer Vision News What is CRF ? What is MRF ? And what are the differences between them? ● A CRF can be thought of as an extension of logistic regression, CRF models conditional probability P(Y|X). For example the code for implementing SVM as a special case of CRF, can be found at this link : ● MRF models the joint probability of both Y and X together. It models P(Y,X), and also can be used to compute P(Y|X=x) for a given input x. The advantage of CRFs is their focus on the “standard” inference problem P(Y|X=x), making them often more precise. On the other hand, that is the only problem they are capable of solving. The advantage of MRFs is that they are completely general, and thus able to model arbitrary inference problems. For instance, let’s assume for some reason some of input x is missing, an MRF can fill-in the missing values because it produces the entire probability distribution. What is Mean-Field Approximation (MFA) ? Approximating the inference and learning problem, using independence assumptions and decomposition into several products, leads to the idea of "mean-field" approximation. In other words, mean-field approximation is a way to simplify the Bayes procedure. MFA can be computed using coordinate ascent. See more at this link .