Rafia Rahim

Since July 2019
Research Assistant at the Department of Cognitive Systems, University of Tübingen
2017 - 2019
Lecturer at the National University of Computer and Emerging Sciences
2014 - 2017
Master of Science in Computer Science at the National University of Computer and Emerging Sciences

During this time I was also employeed as full time Lab Instructor at the same University.

2010 - 2014
Bachelor of Science in Computer Science at the National University of Computer and Emerging Sciences

I am working as a PhD student under the supervision of Prof. Dr. Andreas Zell at the chair of Cognitive System at the Eberhard Karls University of Tübingen. Currently, I am working on DeepStereoVision (a project funded by the German Ministry of Education and Research (BMBF)). The goal of this project is to build efficient deep learning algorithms for stereo vision. Specifically our main goal is to upgrade the legacy stereo vision algorithms (currently used by industrial partner Nerian) with more performant deep learning ones.

Research Interests

  • Computer Vision
  • Deep Learning

Current Projects

In this project, my main focus is to optimize deep neural networks based stereo methods to make them practically deployable in low-end devices. Precisely speaking although current state of the art deep learning based 3D stereo networks give superior performance compared to 2D stereo networks and conventional stereo methods they are not usable in production. So our goal here the goal is to optimize these method for practical deployments without sacrificing their performance.

To this end we have made following progress:

In our recent work, we empirically showed that 3D convolutions act as a major bottleneck in stereo networks. We proposed a set of “plug-&-run” separable convolutions to reduce the computational load (parameters and operations). When evaluated, these convolutions showed reduction in computational load without compromising their performance.

In another work, named MobileStereoNet we have introduced a novel volume construction method to make up for the performance compromise introduced by use of these separable convolutions. We have extended our implementation to both 2D and 3D convolutions based stereo networks. Experiments show that the proposed 2D/3D networks effectively reduce the computational cost without compromising the performance.

Teaching

Topics for Bachelor/Master Theses

We are offering following topics for Bachelor/Master Theses:

  • Bachelor: Implementation of Quantization Algorithms for Model Compression
  • Bachelor: Knowledge Distillation for the training of Lean Student Stereo Network
  • Master: Self Supervised Learning for Deep Stereo Vision

You can find more information at Bachelor Theses and Master Theses pages. If you want to discuss the offered topics or topics of your interest drop me an email at rafia.rahimspam prevention@uni-tuebingen.de

Publications

[1] Rafia Rahim, Faranak Shamsafar, and Andreas Zell. “Separable Convolutions for Optimizing 3D Stereo Networks”. In IEEE International Conference on Image Processing (ICIP), pages 3208-3212, Anchorage, AK, USA, September 2021. [  Link ] [Code] [Poster]
[2] Faranak Shamsafar*, Samuel Woerz*, Rafia Rahim, and Andreas Zell. “MobileStereoNet: Towards Lightweight Deep Networks for Stereo Matching”, In IEEE Winter Conference on Applications of Computer Vision (WACV), Hawaii, USA, January 2022. [ Link | Code ] (*equal contribution)
[3] Muneeb Aadil, Rafia Rahim, and Sibt ul Hussain. "Improving Super Resolution Methods Via Incremental Residual Learning." 2019 IEEE International Conference on Image Processing (ICIP). IEEE, 2019. [Link]
[4] R Timofte, S Gu, J Wu, L Van Gool et al. Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR) workshops, 852-863, 2018. [Link]
[5] Atique ur Rehman, Rafia Rahim, Shahroz Nadeem and Sibt ul Hussain. "End-to-end trained cnn encoder-decoder networks for image steganography." In Proceedings of the European Conference on Computer Vision (ECCV) Workshops 2018. [Link]