Darijan Marcetic

Two-stage cascade model for unconstrained face detection

In this paper, we propose a two-stage model for unconstrained face detection. The first stage is based on the normalized pixel difference (NPD) method, and the second stage uses the deformable part model (DPM) method. The NPD method applied to in the wild image datasets outputs the unbalanced ratio of false positive to false negative face detection when the main goal is to achieve minimal false negative face detection. In this case, false positive face detection is typically an order of magnitude higher. The result of the NPD-based detector is forwarded to the DPM-based detector in order to reduce the number of false positive detections. In this paper, we compare the results obtained by the NPD and DPM methods on the one hand, and the proposed two-stage model on the other. The preliminary experimental results on the Annotated Faces in the Wild (AFW) and the Face Detection Dataset and Benchmark (FDDB) show that the two-stage model significantly reduces false positive detections while simultaneously the number of false negative detections is increased by only a few.

An experimental tattoo de-identification system for privacy protection in still images

An experimental tattoo de-identification system for privacy protection in still images is described in the paper. The system consists of the following modules: skin detection, region of interest detection, feature extraction, tattoo database, matching, tattoo detection, skin swapping, and quality evaluation. Two methods for tattoo localization are presented. The first is a simple ad-hoc method based only on skin colour. The second is based on skin colour, texture and SIFT features. The appearance of each tattoo area is de-identified in such a way that its skin colour and skin texture are similar to the surrounding skin area. Experimental results for still images in which tattoo location, distance, size, illumination, and motion blur have large variability are presented. The system is subjectively evaluated based on the results of tattoo localization, the level of privacy protection and the naturalness of the de-identified still images. The level of privacy protection is estimated based on the quality of the removal of the tattoo appearance and the concealment of its location.

SIFT vs. FREAK: Assessing the usefulness of two keypoint descriptors for 3D face verification

Many techniques in the area of 3D face recognition rely on local descriptors to characterize the surface-shape information around points of interest (or keypoints) in the 3D images. Despite the fact that a lot of advancements have been made in the area of keypoint descriptors over the last years, the literature on 3D-face recognition for the most part still focuses on established descriptors, such as SIFT and SURF, and largely neglects more recent descriptors, such as the FREAK descriptor. In this paper we try to bridge this gap and assess the usefulness of the FREAK descriptor for the task for 3D face recognition. Of particular interest to us is a direct comparison of the FREAK and SIFT descriptors within a simple verification framework. To evaluate our framework with the two descriptors, we conduct 3D face recognition experiments on the challenging FRGCv2 and UMB-DB databases and show that the FREAK descriptor ensures a very competitive verification performance when compared to the SIFT descriptor, but at a fraction of the computational cost. Our results indicate that the FREAK descriptor is a viable alternative to the SIFT descriptor for the problem of 3D face verification and due to its binary nature is particularly useful for real-time-recognition systems and verification techniques for low-resource devices such as mobile phones, tablets and alike.

Real-time face tracking under long-term full occlusions

The identified weaknesses of most of state-of-the-art trackers are inability to cope with long-term full occlusions, abrupt motion, detecting and tracking a reappeared target. In this paper, we present a robust real-time single face tracking system with several new key features: semi-automatic target tracking initialization based on a robust face detector, an effective target loss estimation based on a response of a position correlation filter, a candidate image patch selection for re-initialization supported with a short- and long-term memories (STM and LTM). These memories are used for tracking re-initialization during online learning procedure. The STM is used to select an image patch as candidate for re-tracking based on stored position correlation filters (from current frame) in case of short-term full occlusions, while the LTM stores aggregated position correlation filters (online learned) is used to recover the tracker from long-term full occlusions. Validation of the tracking system was performed by evaluation on a subset of videos from Online Tracking Benchmark (OTB) dataset and our own video.

Deformable part-based robust face detection under occlusion by using face decomposition into face components

In this paper, we propose modifications of deformable part-based models in order to increase the robustness of face detection under occlusion. The modifications are: i) the tree, representing the deformable part-based model of the frontal face, which is partitioned into 11 subtrees representing face components; ii) the weight of each face component which is obtained based on the results of psychological experiments; iii) the introduction of new scoring functions and thresholds; and iv) a new procedure for robust face detection based on the valuation of scoring functions and thresholds. The experiment was performed only for frontal face images, and thus this work is used only as a proof of concept. Based on the encouraging experimental results, we conclude that the proposed method is suitable for extension to detect faces with different poses under occlusions.

Deep learning architectures for tattoo detection and de-identification

The widespread use of video recording devices to obtain recordings of people in various scenarios makes the problem of privacy protection increasingly important. Consequently, there is an increased interest in developing methods for de-identification, i.e. removing personally identifying features from publicly available or stored data. Most of related work focuses on de-identifying hard biometric identifiers such as faces. We address the problem of detection and de-identification of soft biometric identifiers - tattoos. We use a deep convolutional neural network to discriminate between tattoo and non-tattoo image patches, group the patches into blobs, and propose the de-identifying method based on replacing the color of pixels inside the tattoo blob area with a values obtained by interpolation of the surrounding skin color. Experimental evaluation on the contributed dataset indicates the proposed method can be useful in a soft biometric de-identification scenario.

Hybrid Cascade Model for Face Detection in the Wild Based on Normalized Pixel Difference and a Deep Convolutional Neural Network

The main precondition for applications such as face recognition and face de-identification for privacy protection is efficient face detection in real scenes. In this paper, we propose a hybrid cascade model for face detection in the wild. The cascaded two-stage model is based on the fast normalized pixel difference (NPD) detector at the first stage, and a deep convolutional neural network (CNN) at the second stage. The outputs of the NPD detector are characterized by a very small number of false negative (FN) and a much higher number of false positive face (FP) detections. The FP detections are typically an order of magnitude higher than the FN ones. This very high number of FPs has a negative impact on recognition and/or de-identification processing time and on the naturalness of the de-identified images. To reduce the large number of FP face detections, a CNN is used at the second stage. The CNN is applied only on vague face region candidates obtained by the NPD detector that have an NPD score in the interval between two experimentally determined thresholds. The experimental results on the Annotated Faces in the Wild (AFW) test set and the Face Detection Dataset and Benchmark (FDDB) show that the hybrid cascade model significantly reduces the number of FP detections while the number of FN detections are only slightly increased.

Robust initialization for reasoning procedures in a hierarchical heterogeneous knowledge-base

This paper describes a model of a hierarchical, heterogeneous knowledge-base. The proposed model consists of an associative level that is implemented by a Kanerva-like sparse distributed memory (SDM) and a semantic level realized by a knowledge-representation scheme based on the Fuzzy Petri Net theory. The levels are interconnected with forward and backward connections that are used for the robust initialization of multi-reasoning procedures (inheritance, recognition and intersection search) at the semantic level. Multi-reasoning supports reasoning for an unknown concept (i.e., a concept that is not defined at the semantic level), parallel reasoning for more than one concept that is obtained by forward connections from the associative level and used for multiple initialization, and the chaining of associative information retrieval and the reasoning process using forward and backward connections. An example of the initialization of the multi-inheritance procedure is given.