Citation Context

...ocused on the constrained setting of butterfly field guides and uses hand coded visual features specific to that setting, language templates, and predefined attribute lists (lists of color terms etc) =-=[11]-=- to obtain visual representations from4 Tamara L. Berg, Alexander C. Berg, and Jonathan Shih Fig.2. Example input data (images and associated textual descriptions). Notice that the textual descriptio...

Citation Context

...d discriminative power on fine-grained recognition tasks. We demonstrate our approach through systematic experiments on two challenging datasets: CaltechUCSD Birds-200-2011 [22] and Leeds Butterflies =-=[24]-=-. We find on these datasets that our discovered local attributes outperform those generated by human experts and by other strong baselines, on fine-grained image classification tasks. 2. Related work ...

Citation Context

...other. [8, 32, 33, 35] in different ways attempt to find parts of the image that are discriminative without explicit part labels, but cannot achieve the accuracy of a supervised part-based approach. [4, 26] propose interactive approaches which include the system requesting the location of the most discriminative parts from the user. [9] defines a set of just 9546 two coarse parts (the head and body) used to align the images, but do not use fine-scale part locations to define their features. Additional fine-grained recognition work has been demonstrated on trees [13], flowers [19], butterflies [28, 8], and dogs [16, 21, 23]. Although much recognition work continues to use fixed features such as SIFT, HOG, and LBP, there is also work which, like ours, attempts to some degree to learn the feature from a dataset. One branch of this is the work on attributes. [14] and [15] train attribute classifiers based on a set of classes with labeled attributes, then apply the attribute classifiers to novel classes, in the domains of faces and animal species respectively. In both cases, unlike our POOFs, the attributes are chosen by hand. [8] comes closer to our work, automatically finding discriminative ...

Citation Context

...arning of classes and attributes has been explored using Multiple Instance Learning [27] and latent SVMs [29]. Automated discovery of attributes from text and associated images has also been explored =-=[14, 1, 28]-=-. The key advantage of our method is that our parts implicitly model the pose and camera view, which we believe results in more powerful discrimination capabilities. A second line of work has focused ...

Citation Context

...is whether this massive quantity of weakly annotated data can be used to perform metric learning. Although weak forms of supervision have been considered for a number of computer vision related tasks =-=[11, 12]-=-, there is currently more work on metric learning from semisupervised settings [2, 13] than from noisy and weak supervision [14]. In this paper, we focus on a particular form of weak supervision where...

Citation Context

...ct of our approach and are used to drive sentence generation. 2.3 Attributes There is a great deal of ongoing research on estimating attributes for use in computer vision [6], [13], [19], [27], [29], =-=[43]-=- that maps well to our process of estimating modifiers for objects in images. We use the low level features from Farhadi et al. [13] for modifier estimation. Most past approaches have either construct...

Citation Context

...heir structure. Examples include associating names mentioned in the captions to faces depicted in news images (e.g., [23], [24]), verbs to body poses [25], and learning models for recognizing objects =-=[26]-=- and their relative importance [27]. Within natural language processing, most previous efforts have focused on generating captions to accompany complex graphical presentations such as pie charts and b...