Cumulative Attribute Space for Age and Crowd Density Estimation

Abstract

Abstract

Fine-Grained Visual Comparisons with Local Learning

“…This is exactly what helps accuracy for subtle, fine-grained comparisons, and, to some extent, mitigates the impact of inconsistent training comparisons. For an application requiring a full ordering of many images, one could feed our predictions to a rank aggregation technique [10], or apply a second layer of learning to normalize them, as in [9,11,23]. …”

“…At the same time, by not insisting on a single global function to relate all images, we mitigate the impact of inconsistencies in visual comparisons. To our knowledge, no prior work specifically explores fine-grained visual comparisons, and all prior methods assume a single global function is sufficient [9,11,21,23,28]. Furthermore, while local learning methods have been explored for classification [2,6,17,31,33] and information retrieval problems [3,13,16,24], our idea for learning local ranking functions with comparisons is new.…”

“…In [11], rankers trained for each feature descriptor (color, shape, texture) are combined to produce a single global ranking function. Aside from learning to rank formulations, researchers have applied the Elo rating system for biometrics [29], and regression over "cumulative attributes" for age and crowd density estimation [9]. All the prior methods produce a single global function for each attribute, whereas we propose to learn local functions tailored to the comparison at hand.…”

From Semi-supervised to Transfer Counting of Crowds

“…Crowd counting: Various approaches to crowd counting have been proposed [21], including counting-bydetection [20,39,12], counting-by-clustering [6,29], and counting-by-regression [9,10,19,7]. The latter is favoured by most recent studies due to its robustness against occlusion.…”

“…Counting in crowded public spaces is non-trivial due to severe inter-object occlusion, scene perspective distortion, and visual ambiguity introduced by challenging lighting condition and complex human activities. State-of-the-art methods [9,10,19,7] typically adopt regression-based techniques to learn a mapping between low-level features and people count, so as to circumvent explicit object segmentation and detection in crowded scenes. However, these techniques generally require exhaustive frame-wise labelling or even exact head-position annotations [19] to train a regression model.…”