Fast and Accurate Hand Shape Classification

Abstract

Abstract. The problem of hand shape classification is challenging since a hand is characterized by a large number of degrees of freedom. Numerous shape descriptors have been proposed and applied over the years to estimate and classify hand poses in reasonable time. In this paper we discuss our parallel framework for real-time hand shape classification applicable in real-time applications. We show how the number of gallery images influences the classification accuracy and execution time of the parallel algorithm. We present the speedup and efficiency analyses that prove the efficacy of the parallel implementation. Noteworthy, different methods can be used at each step of our parallel framework. Here, we combine the shape contexts with the appearance-based techniques to enhance the robustness of the algorithm and to increase the classification score. An extensive experimental study proves the superiority of the proposed approach over existing state-of-the-art methods.

Real-Time People Counting from Depth Images

“…In order to compare various algorithm variants, we executed each variant on the entire dataset. The results (cumulated for all sequences) are given in table 3: In and Out denote the number of passengers coming in/out the bus determined by the corresponding algorithm variant (1)(2)(3)(4)(5)(6)(7)(8)(9)(10)(11)(12)(13)(14)(15)(16), r In = In/GT In , and r Out = Out/GT Out are the ratios of In/Out and the ground-truth values, and δ denotes the error given as…”

Hand landmarks detection and localization in color images

“…Furthermore, we want to localize all the knuckles along the maxima paths that are currently used to detect the digits. Finally, we intent to take advantage of the extracted information in our shape-based gesture recognition system [36,37]. …”

“…Among the methods for estimating a hand pose, there are solutions based on localizing hand landmarks [6,17,45,51,54], extracting hand shape features [36,37,56], or fitting the parameters of a 3D hand model [15,49,59]. In the last case, subspace learning (linear [15] or non-linear [59]) is applied so as to improve the searching process in a large database of hand images obtained from the model.…”

“…There are mainly two general approaches towards extracting the features [3,41]: (i) appearance-based and (ii) model-based. The former methods involve analysis of the hand silhouette shape [36,37,56] and detecting the landmarks [6,17,45,51,54], which is the main scope of the work reported here and are given more attention in the subsequent section. The model-based methods [15,49,59] are focused on fitting a predefined 3D hand model to an image subject to the analysis-the 3D model parameters form the feature vector which may be further processed.…”

“…Our data sets (HGR1 and HGR2) contain images acquired with controlled and uncontrolled background, which makes skin segmentation challenging [21,22]. In the database, we provide ground-truth SPMs-this makes it possible to validate skin segmentation algorithms, and also enables evaluating hand shape recognition [36] as well as landmark detection and localization relying on the ground-truth data, so as to prevent error propagation. Therefore, all of the results reported in this paper (for each method from Table 1) have been obtained using ground-truth SPMs.…”