HOG active appearance models

Abstract

We propose the combination of dense Histogram of Oriented Gradients (HOG) features with Active Appearance Models (AAMs). We employ the efficient Inverse Compositional optimization technique and show results for the task of face fitting. By taking advantage of the descriptive characteristics of HOG features, we build robust and accurate AAMs that generalize well to unseen faces with illumination, identity, pose and occlusion variations. Our experiments on challenging in-the-wild databases show that HOG AAMs significantly outperfrom current state-of-the-art results of discriminative methods trained on larger databases.

Feature-Based Lucas–Kanade and Active Appearance Models

“…We adopt the concept of highly-descriptive, densely-sampled features within the IC optimization and utilize multi-channel warping at each iteration of the IC optimization which does not greatly increase the computational complexity but significantly improves the fitting performance and robustness. In our previous work [32], we showed that the combination of AAMs with HOG features results in a powerful model with excellent performance. Herein, we apply the above concept for both LK and AAMs by using a great variety of widely-used features, such as Histograms of Oriented Gradients (HOG) [33], Scale-Invariant Feature Transform (SIFT) [34], Image Gradient Orientation kernel (IGO) [15], [20], Edge Structure (ES) [22], Local Binary Patterns (LBP) [35]- [37] with variations [38], and Gabor filters [39]- [41].…”

A survey on face detection in the wild: Past, present and future

“…As discussed above, the problem with fully connected shape models is that they are not easy to solve for the global optima. Certainly, the detection results produced by a tree-based DPM can be used as good initializations for Gauss-Newton [177,178] or regression methods [179].…”

“…For example, current state-of-the-art landmark localization algorithms are usually based on variations of Active Appearance Models that use robust features such as IGOs, HoGs etc. [177,178] or cascade regression schemes based on SIFT or HoG features [179]. In all cases holistic texture models are used by assembling feature vectors that concatenate all facial texture features (i.e., they do not treat each part separately).…”

“…In all cases holistic texture models are used by assembling feature vectors that concatenate all facial texture features (i.e., they do not treat each part separately). Furthermore, in the AAM variations [177,178], the shape is modelled using a fully connected precision matrix. In contrast, DPMs use part-based texture models and a flexible tree-based model, which may lead to unnatural face deformations.…”

A 3D Morphable Model Learnt from 10,000 Faces

“…Using the texture information of the 3D mesh, each virtual camera, which has a known perspective projection matrix, records a realistic synthetic face image with a fixed pose. Therefore, we are able to apply an AAM-based state-of-theart image landmark localisation technique [5], trained for this specific pose and initialised from a state-of-the-art face detector [22,1]. In this way, a set of 68 sparse annotations in the corresponding synthetic view is robustly located and then back-projected on the 3D facial mesh.…”