Mira Slavcheva

Contact Miroslava Slavcheva Email: mira.slavcheva[at]tum.de Address: Technische Universität München Fakultät für Informatik Boltzmannstr. 3 85748 Garching b. München, Deutschland Siemens AG Research & Technology Center Otto-Hahn-Ring 6 81739 München, Deutschland

Research Interests Rigid & Deformable 3D Reconstruction Optimization RGB-D Sensors

About I am a PhD student at the Chair for Computer Aided Medical Procedures of TUM and Siemens AG under the supervision of PD Dr. Slobodan Ilic since March 2015. Before that I obtained an MSc degree in Computational Science and Engineering from TUM and a BSc in Computer Science from Jacobs University Bremen. My current research focuses on rigid and non-rigid 3D reconstruction with RGB-D sensors. In my free time I co-organize the Computer Vision and Medical Image Analysis Meetup with Fausto Milletari, and go rowing with the university team.

News

• [02.2020] The final project from my PhD has just been accepted as a TPAMI journal. We applied recent accelerated optimization theory to our TSDF evolution as a faster converging alternative to KillingFusion^? and SobolevFusion^?.
• [10.2018] PhD defense done! ☺
• [02.2018] Our paper "SobolevFusion: 3D Reconstruction of Scenes Undergoing Free Non-rigid Motion" got accepted as a spotlight presentation at CVPR 2018.
• [10.2017] Our implicit correspondence work got the best paper award of the PeopleCap workshop at ICCV 2017 ☺
• [03.2017] Our paper "KillingFusion: Non-rigid 3D Reconstruction without Correspondences" got accepted as a spotlight presentation at CVPR 2017.
• [09.2016] We are going to demonstrate a 10 FPS version of the Siemens depth sensor used for the 3D-Printed RGB-D Object Dataset in SDF-2-SDF at ECCV 2016 (afternoon session on Thursday, 13. October).

Publications 2020 M. Slavcheva, M. Baust, S. Ilic Variational Level Set Evolution for Non-rigid 3D Reconstruction from a Single Depth Camera IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI). The final publication is available at ieeexplore.ieee.org. (supplementary) (bib) 2018 C. Kozlov, M. Slavcheva, S. Ilic Patch-based Non-rigid 3D Reconstruction from a Single Depth Stream International Conference on 3D Vision (3DV), Verona, Italy, September 2018 (bib) M. Slavcheva, M. Baust, S. Ilic SobolevFusion: 3D Reconstruction of Scenes Undergoing Free Non-rigid Motion IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Salt Lake City, USA, June 2018 [spotlight] (supplementary) (bib) 2017 M. Slavcheva, W. Kehl, N. Navab, S. Ilic SDF-2-SDF Registration for Real-time 3D Reconstruction from RGB-D Data International Journal of Computer Vision (IJCV), Springer Verlag. The final publication is available at www.springerlink.com. (bib) M. Slavcheva, M. Baust, S. Ilic Towards Implicit Correspondence in Signed Distance Field Evolution PeopleCap Workshop, IEEE International Conference on Computer Vision (ICCVW), Venice, Italy, October 2017 (bib) M. Slavcheva, M. Baust, D. Cremers, S. Ilic KillingFusion: Non-rigid 3D Reconstruction without Correspondences IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Honolulu, USA, July 2017 [spotlight] (supplementary) (bib) 2016 M. Slavcheva, S. Ilic SDF-TAR: Parallel Tracking and Refinement in RGB-D Data using Volumetric Registration British Machine Vision Conference (BMVC), York, UK, September 2016 (abstract) (bib) M. Slavcheva, W. Kehl, N. Navab, S. Ilic SDF-2-SDF: Highly Accurate 3D Object Reconstruction European Conference on Computer Vision (ECCV), Amsterdam, The Netherlands, October 2016 (supplementary) (bib) Teaching Assistance Lecture: Tracking and Detection in Computer Vision (WS 16/17) Hauptseminar: Recent Trends in 3D Computer Vision (WS 16/17) Lecture: Tracking and Detection in Computer Vision (WS 15/16) Hauptseminar: Recent Trends in 3D Computer Vision (WS 15/16) Lecture: Tracking and Detection in Computer Vision (WS 14/15)

Datasets The single-stream RGB-D sequences with ground-truth canonical pose 3D models we acquired for our KillingFusion? paper are available here. The 3D-Printed Dataset introduced in the SDF-2-SDF paper is available here. It contains groundtruth CAD models and camera trajectories for 5 objects, each of which is scanned both with turntable and handheld motion. We provide synthetic, industrial- and Kinect-quality RGB-D sequences. Projects Implicit-to-Implicit Registration for Rigid-Body Motion Estimation This is a method for precise registration from range data, in which pairs of signed distance fields are aligned by means of minimization of their direct per-voxel difference. Some of the advantages over cloud registration techniques are the absence of correspondence search and the denser formulation, which lead to higher accuracy and a larger convergence basin. It is used as the basis for SDF-2-SDF, an object reconstruction pipeline that combines frame-to-frame camera tracking and multiview pose optimization, and SDF-TAR, an extension to larger scenes and SLAM. The project started with my master thesis (Unified Pipeline for 3D Reconstruction from RGB-D Images using Coloured Truncated Signed Distance Fields), supervised by Wadim Kehl. It was selected for presentation at the Young Researcher Forum at GCPR 2015. Later, I got a chance to present the underlying approach as an oral at the WiCV workshop at CVPR 2016 (poster). Signed Distance Field Evolution for Non-rigid 3D Reconstruction This is a variational framework for 3D reconstruction of non-rigidly moving surfaces captured with a single RGB-D camera. We estimate a dense deformation field that warps a given SDF to a target SDF, steered by a data term that enforces voxel-wise alignment. To ensure plausible shapes and impose regularity over the field, we propose different strategies, such as the approximately Killing vector field regularizer used in KillingFusion, and the gradient flow in Sobolev space used in SobolevFusion. The advantage of these SDF-based evolution schemes is their ability to handle topological changes and rapid motions, which are challenging for mesh-based techniques. However, this entails loss of data association. To recover correspondences, we study different possibilities with the lowest-frequency eigenfunctions of the Laplacian matrix representing the SDFs. For moderate motions, we are able to estimate data association implicity via a Laplacian eigencolour term in the evolution energy. For larger motions it is not sufficient, so we explicitly determine correspondences via signature matching of lower-dimensional embeddings of the eigenfunctions. Although the estimated correspondences are not always as reliable as those yielded by approaches that rely on correspondences for the deformation field calculation, we believe that this is a neat way to incorporate data association in a framework that is able to handle changing topology.