Research Topics
Object Recognition and Feature Matching
During my PHD, the first problem that I addressed was Feature matching. Addressing such an interesting problem with very competitive approaches was not an easy task. It took time to understand the literature and find the potential improvements over state-of-the-art approaches. The tradeoff between accuracy and complexity was a big issue. The way I addressed this problem was to look for some novel objective function for feature matching while avoiding solving a quadratic assignment problem.
Next is my work on Object recognition. In this work, we managed to find some embedding space for feature points coming from different images. Interestingly the embedded space enabled us to define a similarity measure between pairs of images based on Hausdorff distance. This measure was used to train an SVM for classification with very competitive rates with state-of-the-art.
Next is my work on Pose estimation . In this work a different problem was addressed. The underlying approach was built on standard kernel regression. However, the key contribution was to learn a conceptual manifold from local features extracted from object images. Most of manifold learning methods rely on holistic representation of the object image. In our work we showed that we can build conceptual manifold for object pose using local features.
Other works during PHD and recently after are In summary the work in my PHD was centered on object recognition problems. The solutions we found were mainly about manifold learning from local features.

Activity and Gesture Recognition
In 2012 our research group was granted a fund from Microsoft ATL(Cairo) to pursue research in Activity and gesture Recognition from skeletal data. We successfully contributed two descriptors of skeletal data . Now these descriptors are recognized among the community later in CVPR14, ECCV14 and BMVC14. We made a follow up on the role of feature selection on large descriptors in. This work purpose was to accurately detect the action points. I am Co-advising a PHD student with Professor Lynn Abbott at Virginia Tech. She is working on facial expression recognition from Kinect data. Her work combines basic 2D image features with 3D features extracted from the mesh of the face. The landmarks are returned by the Kinect sensor.

Assistive Computer Vision
While supervising undergraduate students during their graduation projects, some of them were working on assistive computer vision problems like two graduation projects for Sign Language Recognition for deaf and mute people. Also three other graduation projects were to assist blind and visually impaired people. Also, two High school students under my supervision were working on LipReading and their work was shortlisted among the top 90 projects by google science fair 2014.

Contour Detection and Image Segmentation
My M. Sc. Student  Ahmed Taha is working on Image segmentation. We use an approximation for Laplacian eigenvectors that relaxes the actual graph embedding solution. The results for interactive image segmentation using scribble annotation exceeded state-of-the-art algorithm like Boykov and Jolly as well as many other algorithms. The work is accepted in ICIP15.

Manifold/ Machine Learning More or less, the work on object recognition I did on my PHD was based on manifold learning and I am keeping my interest on that. Also the recent work that is about contour detection and interactive image segmentation is based on the graph representation of the manifolds. I worked on different standard machine learning problems including Kernel regression in and unsupervised manifold embedding, manifold learning with supervised learning using SVM in, graph based semi-supervised learning in. In these works I used kernel methods.

Reading Group(2013/2014)

Dr. Marwan Torki and Dr. Mohamed Hussein are leading a group reading meetings every week. Presenters includes Mohamed Gowayyed, Amr Sharaf, Moustafa Meshry, Mahmoud Fayyaz, Ahmed Mamdouh, Ahmed Taha, Mohamed Othman, KarimYaser, Ahmed ELket, Ahmed Eltayebany and Ahmed Mesbah. Attendence is open every Tuesday from 10:30 to 11:30

Date Presentation Title Resources Presenter
29-10-13 Group Sparsity and Geometry Constrained Dictionary Learning for Action Recognition from Depth Maps (ICCV13) Paper, Presentation Marwan Torki
12-11-13 WORK for CVPR14   Amr Sharaf
26-11-13 Real-Time Human Pose Recognition in Parts from a Single Depth Image (CVPR11) Paper, Presentation Mohamed Gowayyed
3-12-13 Learning the Manifolds of Local Features and Their Spatial Arrangments (CVPR10,ICCV11) Thesis , Presentation Marwan Torki
10-12-13 Learning the Manifolds of Local Features and Their Spatial Arrangments (CVPR10)   Marwan Torki
17-12-13 Learning the Manifolds of Local Features and Their Spatial Arrangments (CVPR10)   Marwan Torki
24-12-13 Learning the Manifolds of Local Features and Their Spatial Arrangments (ICCV11)   Marwan Torki
  31-12-13 till 5-2-14 Break for Final Exams and grading period    
6-2-14 Efficient Subwindow Search (ESS) Paper,Presentation Moustafa Meshry
13-2-14 Semi-Supervised Learning and Relations to Graph cut Presentation Ahmed Taha
20-2-14 Novel System for Deaf and Mute Using Lipreading (Best CS project in EGCSEF 2014) Presentation Ahmed Eltayebany and Ahmed Mesbah
27-2-14 Sketch Tokens Paper Mohamed Gowayyed
6-3-14 Oprical Flow (Obstacle Avoidance / Motion segementation/ Ground Plane Detection) Presentation Mohamed Othman, Ahmed Elket, Karim Yaser, Ahmed Magdy
13-3-14 Online Motion Segmentation uisng Dynamic Label Propagations (ICCV13) Paper, Defense Presentation Marwan Torki
20-3-14     Amr Sharaf
3-4-14     Mohamed Hussein