Ejaz Ahmed

Fifth year PhD Student
Computer Sience
University of Maryland - College Park

3364, A.V. Williams
College Park - 20740
Maryland, USA

About Me

I am a fifth year PhD student in the Department of Computer Science at University of Maryland, College Park. I am also a member of Institute for Advanced Computer Studies (UMIACS) and Computer Vision Laboratory. I am working in the area of computer vision, pattern recognition and machine learning with Prof. Larry Davis as my advisor. I recieved MS in computer science from University of Maryland, College Park in May 2013. During my PhD I have interned at various places in academia and industry such as, INRIA-Grenoble (France), Adobe Research in San Jose, TTI Chicago and Mitsubishi Electric Research Lab (MERL) in Cambridge. Before coming to Maryland, I received BTech in computer science from IIIT-Hyderabad, India with an honors degree from Center for Visual Information Technology.

Current Objective

Right now I am seeking a full-time position in the field of computer vision to apply and enhance my research and technical skills.

Projects and Publications

Deep Learning for Person Re-identification

Coming Soon...

This work was done during summer 2014 internship at Mitsubishi Electric Research Laboratories (MERL) with Dr. Michael J. Jones and Dr. Tim K. Marks. We have proposed a deep convolutional architecture for the problem of person re-identification. We show state-of-the-art results on various re-id datasets. More details to be declared soon.
Submitted to CVPR 2015.

Selecting Filters for Object Detection and Beyond

Coming Soon...

This work was done during fall 2013 internship at TTI Chicago with Prof. Greg Shakhnarovich and Prof. Subhransu Maji.
A common way to represent a visual category is to represent it as a collections of filters. The main bottleneck in training such systems is the selection of a subset of good filters from a large number of possible choices. We show that one can learn a universal model of part "goodness" based on properties that can be computed from the filter itself. This allows us to quickly discard filters that are not promising thereby speeding up the training procedure. Applied to training the poselet model, our automated selection procedure allows us to improve its detection performance on the PASCAL VOC data sets, while speeding up training by an order of magnitude. Similar results are reported for exemplar SVMs.
Using Human Knowledge to Judge Part Goodness: Interactive Part Selection
Ejaz Ahmed, Subhransu Maji, Gregory Shakhnarovich, and Larry S. Davis.
Conference on Vision and Pattern Recognition Workshop (CVPRW) 2014.

Knowing a Good HOG Filter When You See It: Efficient Selection of Filters for Detection
Ejaz Ahmed, Gregory Shakhnarovich, and Subhransu Maji.
European Conference on Computer Vision (ECCV) 2014, Zurich (Oral).

paper / supplementary material / talk slides / poster

Semantic Object Selection

Coming Soon...

This work was done during summer 2013 internship at Adobe Research, San Jose with Dr. Scott Cohen and Dr. Brian Price.
Interactive object segmentation has great practical importance in computer vision. Many interactive methods have been proposed utilizing user input in the form of mouse clicks and mouse strokes, and often requiring a lot of user intervention. In this paper, we present a system with a far simpler input method: the user needs only give the name of the desired object. With the tag provided by the user we query a text image database to gather exemplars of the object. Using object proposals and borrowing ideas from image retrieval and object detection, the object is localized in the image. An appearance model generated from the exemplars and the location prior are used in an energy minimization framework to select the object. Our method outperforms the state-of-the-art on existing datasets and on a more challenging dataset we collected.
Semantic Object Selection
Ejaz Ahmed, Scott Cohen, and Brian Price.
Conference on Computer Vision and Pattern Recognition (CVPR) 2014

paper / supplementary material / talk slides / poster / spotlight

Object Detection

Frame 1

I have developed an object detection pipeline based on partial least squares which a class aware dimensionality reduction technique. We have also proposed another linear dimensionality reduction method, Composite Discriminant Factor (CDF) analysis, which searches for a discriminative but compact feature subspace that can be used as input to classifiers that suffer from problems such as multi-collinearity or the curse of dimensionality. The subspace selected by CDF maximizes the performance of the entire classification pipeline, and is chosen from a set of candidate subspaces that are each discriminative.
Composite Discriminant Factor Analysis. (paper)
Vlad I. Morariu, Ejaz Ahmed, Venkataraman Santhanam, David Harwood, and Larry S. Davis.
Winter Conference on Applications of Computer Vision (WACV) 2014.

Real Time Object Tracking
We track multiple objects in videos by first selecting the appropriate feature which discriminates object from background and then applying iterative algorithm to track the objects in subsequent frames. The whole process is done in parallel on GPU(CUDA) which gives us increase in performance (68 fps for 320 X 240 video as compared to 11 fps when done on CPU).
Tracking for Entertainment and Interaction.
Prakhar Jain, Ejaz Ahmed, and Dasari Pavan Kumar.
ACM SIGGRAPH Symposium on Interactive 3D Graphics and Games 2009 (I3D), Boston.
Frame 1

Batch Cut This work attempts to address the problem of object-segmentation from a collection of images.By combining existing image segmentation approaches with simple learning techniques, we attempt to minimize user's involvement in segmentation of images in a group of images. We propose a supervized learning paradigm in which user trains only on a small number of representative images. Based on this training appropiate features are selected and models representing the classes are built. For a new image (image from the batch) energy corresponding to every pixel is calculated and an undirected weighted graph is constructed using these enrgies. These energies are optimized using graph cut.

KinSeg This paper presents an interactive segmentation framework that incorporates both color and depth information of a scene obtained by a Kinect camera. This interactive framework allows users to specify foreground objects and background simply sketching with a brush. These user sketches are then treated as hard constraints. In addition, region and boundary information are used as soft constraints. Based on these constraints, we formulate the segmentation task as an energy minimisation problem. The energy function consists of both the color and depth information of the scene. We use scene knowledge to automatically determine the relative contribution of the depth and color terms to the energy function. Finally, a globally optimal segmentation is obtained by solving this problem using Graph Cuts.

Fast GrabCut Problem of segmentation is computationally intensive and with the increasing resolutions of camera the normal CPU algorithms are turning out to be slow. We present a fast GPU(CUDA) based segmentation method. This method segments the object by first modelling the object foreground and background using Gaussian Mixture Models and then iteratively minimizes the energy using iterative GraphCuts. Both these steps are done by exploiting the computational power of GPU.

Frame 1

Real Time 3D Video Segmentation
In this work the object to be segmented is marked manually in the initial frame of the video. We track the object in subsequent frames. Using this tracking the object and background GMM models are created. Few frames are clubbed together and 3D GraphCuts is applied on them to segment the object. Such 3D GraphCuts ensure coherency and continuity of segmented object across frames. Since this task is computationally intensive and also parallelizable, we are doing this using CUDA GPU.
Real Time Object Segmentation from Videos.
Ejaz Ahmed, Prakhar Jain, and P.J. Narayanan.
Indian Conference for Academic Research by Undergraduate Students 2010 (ICARUS), Kanpur.

Frame 1

GPU Stereo
In this project various stereo algorithm were studied and implemented. Primary focus was given to coarse-to-fine stereo algorithm and its adaptive variants. This project was done as an intern at INRIA Grenoble, France in the summer of 2010. Internship involved understanding of algorithms and their implementation on GPU/CUDA.

Frame 1

Fitting Surfaces on 3 D models using MLS (moving least square) on GPU
The project involves providing a definition of a smooth manifold surface from a set of points close to the original surface. This is done by locally approximating the surface with polynomials using moving least squares (MLS). For fitting curves on these points, k- nearest neighbors are calculated for each point. Then for each point normal is calculated using eigen-values and eigenvectors and finally curve is fitted. Since 3 D model are made up of large number of points (order of Millions), the whole process becomes computationally expensive and highly parallelizable. Thus we exploit the computing ability of the Graphics Processing Unit (GPU) using NVIDIA.s CUDA to speed up the process.

Frame 1

Fast Gist
In this project we have parallelized the GIST feature exploiting the computation power of the GPU with the help of Nvidia.s CUDA.

Frame 1

Work Experience

Teaching Experience (TA)

  • Introduction to Computer Systems (CMSC 216) under Prof. Alan Sussman, Fall'10, at UMD.
  • Artificial Intelligence under Prof. Anoop Namboodiri, Spring'10, at IIIT-H.
  • Computer System Organization under Prof. P.J. Narayanan, during Spring'09, at IIIT-H.
  • Computer Programming under Prof. C.V. Jawahar, during Monsoon'09, at IIIT-H.
  • Maths-I under Prof. C.N. Kaul, Monsoon'08, at IIIT-H.

Graduate Coursework at UMD

  • Linear Subspaces and Manifolds in Computer Vision and Machine Learning, (CMSC 828), Prof. David Jacobs, Spring'13.
  • Image Segmentation, (CMSC 828B), Prof. David Jacobs, Spring'12
  • Computational Geometry, (CMSC 798), Prof. David Mount, Spring'12
  • Statistical Pattern Recognition, (CMSC 828C), Prof. Rama Chellappa, Fall'11
  • High Performance Computing, (CMSC 714), Prof Alan Sussman, Fall'11
  • Computational systems biology and functional genomics, (CMSC 858P), Prof. Héctor Corrada Bravo, Spring'11
  • Database Management Systems(CMSC 724), Prof. Amol Deshpande, Spring'11
  • Image Understanding, (CMSC 828P), Prof Rama, Chellappa, Fall'10
  • Machine Learning, (CMSC 726), Prof. Lise Getoor, Fall'10