Learning to detect unseen object classes by between-class attribute transfer

On-the-fly feature importance mining for person re-identification

“…Second, the conventional top-down approach imposes weights on certain feature types that are considered optimal in a universal sense; while the bottom-up approach aims to discover a set of discriminative features and quantify their importance specific to each individual. From another perspective, the notion of bottom-up learning can also be interpreted as a process of unsupervised discovering latent attribute (see Section 3.1), which is largely different from existing top-down supervised attribute learning [16,15] that requires exhaustive humanspecified attributes.…”

Transductive Multi-view Embedding for Zero-Shot Recognition and Annotation

“…In particular, while both visual attributes [18,6,21,11] and linguistic semantic representations such as word vectors [23,7,34] have been independently exploited successfully, it remains unattempted and not straightforward to exploit synergistically multiple semantic 'views'. This is because they are of very different dimensions and types and each suffers from different domain shift effects discussed above.…”

“…Such a semantic representation is assumed to be shared between the auxiliary and target datasets. More specifically, apart from class label, each auxiliary data point is labelled by a semantic representation such as visual attributes [18,6,21,11], semantic word vectors [23,7,34] or others [28]. A projection function mapping low-level features to the semantic space is learned from the auxiliary dataset by either classification or regression models.…”

“…In particular, a single 'prototype' of each target class is specified in the semantic space. Depending on the semantic space, this prototype can be an attribute annotation vector [18] or a word vector inferred from the target class name [7].…”

Interestingness Prediction by Robust Learning to Rank

“…[28] focuses on key frames so essentially treats it as an image interestingness problem, whilst [21] is the first work that proposes benchmark video interestingness datasets and evaluates different features for video interestingness prediction. In a broader sense of attributes [26,11,12,27,13] interestingness can be considered as one type of relative attributes [35], although those attributes, such as how smiling a person is, are much less subjective. Computational models of interestingness Most earlier work casts the aesthetics or interestingness prediction problem as a regression problem [22,7,19,28].…”