Distinctive Parts for Relative attributes

Naga Sandeep Ramachandruni (homepage)


Visual Attributes are properties observable in images that have human-designated names ( e.g., smiling, natural) and they are valuable as a new semantic cue in various vision problems like facial verification, object recognition, generating description of unfamiliar objects and to facilitate zero shot transfer learning etc. While most of the work on attributes focuses on binary attributes (indicating the presence or absence of attribute) the notion of relative attributes as introduced by Parikh and Grauman in ICCV 2011 provides an appealing way of comparing two images based on their visual properties than the binary attributes. Relative visual properties are a semantically rich way by which humans describe and compare objects in the world. They are necessary, for instance, to refine an identifying description (the rounder pillow; the same except bluer), or to situate with respect to reference objects (brighter than a candle; dimmer than a flashlight).  Furthermore, they have potential to enhance active and interactive learning, for instance, offering a better guide for a visual search (find me similar shoes, but shinier or refine the retrieved images of downtown Chicago to those taken on sunnier days). For learning relative attributes a ranking svm based formulation was proposed that uses globally represented pairs of annotated images. In this thesis, we extend this idea towards learning relative attributes using local parts that are shared across categories.

First we propose a part based representation that jointly represents a pair of images. For facial attributes, part corresponds to a block around a landmark point detected using a domain specific method. This representation explicitly encodes correspondences among parts, thus better capturing minute differences in parts that make an attribute more prominent in one image than another as compared to global representation. Next we update this part based representation by additionally learning weights corresponding to each part that denote their contribution towards predicting the strength of a given attribute.

We call these weights as significance coefficients of parts. For each attribute the significance coefficients are learned in a discriminative manner simultaneously with a max-margin ranking model. Thus the best parts for predicting relative attribute more smiling will be different from those from predicting more eyes open. We compare the baseline method of Parikh and Grauman with the proposed method under various settings. We have collected a new dataset of 10000 pair wise attribute level annotations using images from labeled faces in the wild ( LFW) dataset particularly focusing on large variety of samples in terms of poses, lightning conditions etc and completely ignoring the category information while collecting attribute annotation . Extensive experiments demonstrate that the new method significantly improves prediction accuracy as compared to the baseline method. Moreover the learned parts also compare favorably with human selected parts, thus indicating the intrinsic capacity of the proposed framework for learning attribute specific semantic parts. Additionally we illustrate the advantage of the
proposed method with interactive image search using relative attribute based feedback.

In this work, we also propose relational attributes, which provide a more natural way of comparing two images based on some given attribute than relative attributes. Relational attributes consider not only the content of a given pair of images, but also take into account its relationship with other pairs, thus making the comparison more robust.


Year of completion:  July 2016
 Advisor : Prof. C.V. Jawahar

Related Publications