SCCS Colloquium - May 15, 2019
|Date:||May 15, 2019|
|Time:||15:00 - 16:00|
Vladimir Poliakov: Inferring 3D Human Pose in Real-Time on Consumer Smartphones: A Lightweight Neural Approach
This is a Master's thesis submission talk. Vladimir is advised by Prof. Hans-Joachim Bungartz.
Human pose estimation from images and videos is a significant research area in computer vision and deep learning. The general trend in HPE is to develop more complex models to achieve higher accuracy. However, in many real-world applications estimation tasks must be performed in a timely manner. The development of effective models designed for mobile devices is still an open question. In this thesis, we present two lightweight three-dimensional human posture estimation models, based on the idea of the stacked hourglass network, a deep convolutional architecture.
We analyze and benchmark a lightweight, optimized unit that substitutes the original heavy unit of the hourglass model. Such a unit, based on the idea of depthwise separable convolution technique, allows us to significantly reduce the number of trainable parameters and the number of floating point operations. We also consider the rectangular format of tensors for a smartphone-made image. This format also allows us to simplify the model without loss of quality.
After the analysis of the optimized block for a two-dimensional model, we proceed to the development of three-dimensional models on its basis. The first lifting approach links the already estimated two-dimensional pose of a person with a three-dimensional in space. In this approach, we used our optimized two-dimensional model to obtain the keypoints of a person in two-dimensional space, and then put these points in three-dimensional, using a conventional neural network with a residual structure. The second weakly supervised approach combines an optimized two-dimensional model with an optimized depth regression module in a single deep neural network. We optimized this solution by offering an optimized block and resolution, reducing the number of parameters and the number of floating point operations.
We present thorough quantitative benchmarking first on a two-dimensional model using the MPII standard, which confirmed the efficiency of using the optimized block and the rectangular format of tensors. To validate designed three-dimensional approaches, we experimented on several datasets. The models are trained on MPII and Human 3.6M and evaluated using the Human 3.6M standard. These experiments show strong performance compared to other state-of-the-art 3D models.
This thesis also provides qualitative experiments for described models on Apple's mobile devices. Thus the results show not only the accuracy of the models but also the ability to use these models on the 4-year-old iPhone model at a high frame rate.
Keywords: human pose estimation, deep learning, computer vision
Qunsheng Huang: Helicopter Simulations with preCICE
This thesis will attempt to couple the CFD/CSD solvers used to simulate helicopter rotor blade dynamics, TAU and CAMRAD, respectively, using the preCICE coupling library. Main areas of work will involve the mapping of node data between the 3D CFD simulation and the 1D CSD simulation while preserving stability of the overall simulation. Other than the possible mapping improvements, the introduction of preCICE is intended to effect various performance improvements while increasing the overall flexibility of the coupled codes.
Aleksei Dolgodvorov: Parallel algorithms for reducing the bandwidth of symmetric matrices.
This a Master's thesis introduction talk. Aleksei is advised by Michael Rippl.
Reduction to tridiagonal form is an essential step in eigenvalue computations for symmetric matrices. This type of matrices with banded structure arises in particular from generalized eigenvalue problems. The algorithm proposed by Crawford is used to transform a generalized eigenvalue problem to a standard form where banded structure of the matrix is preserved. But in case of a large bandwidth, it should be reduced. In the current work the method for symmetric band reduction proposed by Bischof and Lang is used and the parallel implementation of this algorithm is developed.
Keywords: Eigenvalue problem, symmetric matrices, symmetric band reduction, parallel implementation