Associate Professor AJMAL SAEED MIAN
ARC Australian Research Fellow (2011 - 2015)
Computer Science & Software Engineering
The University of Western Australia
35 Stirling Highway, Crawley, WA 6009
Phone: +61-8-64882702    Fax: +61-8-64881089
email: ajmal.mian [shift+2] uwa.edu.au

B.E. Avionics 1993
M.S. Information Security 2003
Ph.D. Computer Science 2006

Google Scholar Citations

YouTube Channel

I lead the Macine Intelligence Group (MIG) and am a member of Vision and Visualization (VisVis) and the Complex Data Modelling Group (CDM).

PhD Scholarships:
 PhD scholarships are available for outstanding students. International students can apply in the main IPRS scholarship rounds and local (Australian and New Zealand) students can apply in the UPA scholarship rounds. The following projects are available but students are welcome to propose their own projects within the scope of computer vision:

1. Hyperspectral image analysis for remote sensing and computer vision.
2. Image set modelling and classification.
3. RGB-D video based object recognition and action classification.
4. 3D morphometric analysis of faces or other objects.
5. Video based 

Send me an email with your 2 page CV (in PDF format) stating your grades and publications. Research proposal is not required at this stage.

Postdoc positions:  Currently not available.

Bayesian Sparse Representation for Hyperspectral Image Super Resolution (CVPR 15)
A high resolution image is fused with a low resolution hyperspectral image using non-parametric Bayesian sparse representation to obtain a high spectral and high spatial resolution image. Bayesian sparse representation is proposed.
Automatic detection of a large number of facial landmarks (CVPR 15)
Level set curves are evolved over the 3D face with adaptive geometric speed functions to extract effective seed points for dense correspondences to be established by minimizing the bending energy between patches around seed points of faces
NKTM: Nonlinear Knowledge Transfer Model for cross-view action recognition (CVPR 15)
NKTM is a deep network, with weight decay and sparsity constraints, which finds a single shared high-level virtual path from videos captured from different unknown viewpoints to the same canonical view. A single NKTM is learned for all actions and all camera viewing directions without action labels. 
Hyperspectral Face Recognition
The most comprehensive study to date on hyperspectral face recognition. We propose two techniques and perform extensive comparisons with existing hyperspectral, RGB, gray scale and image set classification algorithms. We release the largest database in this domain.
Deep ELM for image set classification
Non-linear structure of image sets is learned with Deep Extreme Learning Machines (DELM) that generalize even with limited training data. Domain specific DELMs are learned in an unsupervised fashion and then adapted to class specific representations.
Dense 3D face correspondence
Starting from automatic sparse correspondences, our algorithm triangulates existing correspondences and expands them iteratively along the triangle edges. A deformable model (K3DM) is constructed from the 3D faces and an algorithm is proposed for morphing the K3DM to fit unseen faces. 
Hyperspectral Image Super Resolution
We propose Generalized Simultaneous Orthogonal Matching Persuit with positivity constraint to  obtain a high spectral + spatial resolution hyperspectral image given low spatial high spectral and high spatial resolution RGB images.
HOPC: Histogram of Oriented Principal Components for action recognition
A new descriptor is proposed for pointcloud representation along with a spatio-temporal kepoint detection method. HOPC extracted in a local coordinate frame at keypoints is viewpoint invariant. Code avaiable. 
Direct Fiedler Vector computation for two-way NCut
An algorithm is proposed to directly compute the 2nd last eigenvector (Fiedler Vector) of the normalized graph Laplacian. Hierarchicah two-way graph partitioning is performed and applied to image set classification.
Action classification with Locality Constrained Linear Coding
Depth video is divided into spatiotemporal cells and represented with HOG3D. A dictionary is derived from the descriptors and used to code them with nearby (local) dictionary elements. Classification is performed using logistic regression with L2 regularization.
SUnGP: Greedy sparse approximation for hyprspectral unmixing
Sparse Unmixing via Greedy Persuit is proposed which first identifies a subspace of L endmembers (dictionary atoms) and then prunes it based on non-negativitiy to detect the most likely endmember. The residue is then updated and the process is repeated.
Perceptual differences between men and women
Gender classification is best performed by geodesic and Euclidean distances between biologically significant facial landmarks. More importantly, gender classification with the selected facial features replicate the perceptual gender bias related to facial expressions.
Real-time action recognition with histograms of depth gradients and RDF
Depth histograms and 3D joint space-time motion volumes are extracted and feature selection is performed using Random Decision Forests (RDF). Over 100 fps are processed on a single core 3GHz machine.
Hyperspectral ink mismatch detection
Hyperspectral imaging can be used to detect forgery in documents. We analyze 5 black and 5 blue inks to detect mismatch between inks in the spectral range of 400-720nm (33 bands at 10nm step). Dataset is available.
Repeated constrained sparse coding with partial dictionaries for hyperspectral unmixing
Dictionaries are made from pure spectra and maximally correlated bands (rows of dictionaries) are systematically removed to perform constrained sparse coding of the mixed spectra. The sparse codes provide the relative abundance of materials.
Semi-supervised Spectral Clustering for Image Set Classification
Query set and training sets are co-clustered in an unsupervised way. Class-cluster distribution is used for classification. A fast eigen solver is also proposed which provides more accurate classification in addition to performance gain.
Self regularized non-negative coding and adaptive distance metric learning
Not every point in a linear face subspace is a valid face. We impose these two constraints when fiding the nearest points between image-sets. We also formulate adaptive distance metric learning for image-sets given label consistancy of sets.
ContourCode: Multispectral palmprint encoding and recognition
Based on the non-subsampled contourlet transform the directional response of the palmprint lines is encoded into a hash table that is compact and facilitates fast recognition time. ContourCode outperforms CompCode, ContCode, DogCode, OLOF on two databases.
Sparse Approximated Nearest Points for Image Set Classification
Image sets are represented by the image samples and their affine hull model. The dissimilarity between sets is measured as the distance between their nearest points that can be sparsely approximated from the sample images of the respective set.
Robust Realtime Feature Detection in Raw 3D Face Images
We avoid preprocessing 3D data for noise/spike removal, hole filling and detect the outer eye corners and nose tip with over 99% accuracy using Haar features extracted from 3D image gradients. [download code/haarcascades]
Facial Gender Scoring: Objectivity of Perception
Gender score is a continuum and used in psychosocial studies. We show that human perception of facial gender can be objectively measured using geodesic distances between anthropometric facial landmarks. High correlation with 70 human raters was obtained.
3D Scale Invariant Keypoint Detection
Multi-scale keypoints are detected on 3D objects/ partial surfaces for pose and scale invariant object retrieval from cluttered scenes. Features are ranked using a quality measure and the best scale for feature (at a keypoint) is automatically determined from data.
Desktop Optics for Face Recognition and 3D Reconstruction
Facial features are extracted at multiple scales and orientations and projected separately to linear subspaces for illumination invariant recognition. A model based approach and Support Vector Regression is used to reconstruct 3D faces.
Detection and Recognition of 3D Ears
Haar features are used to detect and extract ears regions. Keypoints are then localized on the 3D ear surface for extracting local features. Features are matched using various geometric constraints for person identification.
Smart Surveillance
Realtime face tracking is done with a single PTZ camera. See demos and code for controlling Sony PTZ network camera.
Video-based Face Recognition
Unsupervised learning is performed from local features for video-based face recognition. An extension of this technique performs online learning (with near realtime feedback) and face recognition.
Shade Face: Muliple Image-based face recognition
A global signature of the face, under varying illumination, is constructed using contourlet coefficients. Illumination is varied using a computer screen. Face signatures are matched using sliding windows and the results are compared to 3D face recognition.
Keypoint based 3D Face Recognition
Unique keypoints are detected on 3D faces to extract invariant features that are projected to the PCA subspace and matched. Similarity between faces is measured using graphs (meshes) constructed from the matching keypoints on individual faces.
Non-Rigid 3D Face Recognition
Expression deformations are modelled in PCA subspace and the model is used to morph out expressions from 3D faces. The resulting faces are used for expression invariant 3D face recognition.
Facial Expression Classification
An unseen 3D face, under any facial expression, is decomposed into an estimated 3D neutral face and expression deformations. This decomposition classifies the facial expression and improves the accuracy of3D face recognition.
3D object recognition in cluttered scenes
An algorithm is prposed for 3D object recognition in cluttered scenes and its performance
is compared to the spim images on the same database. The database can be downloaded for comparisons.
3D Multimodal hybrid face recognition
This research has focused on automatic pose correction, keypoint identification, local feature extraction and a region-based matching for robust 3D face recognition.
Illumination normalization of colour face images
The Phong's lighting model is used to normalize illumination in face images given their 3D models. Both Lambertian and specular components are taken into account. The illumination normalized images give higher face recognition accuracy.
Internet key exchange
A new protocol (Arcanum) has been proposed for secure key exchange over the Internet. Arcanum is secure, robust to DoS attacks and more efficient in terms of CPU consumption and network traffic. Java code can be downloaed.
3D modeling
A representation for 2.5D surfaces is proposed based on third order tensors. Tensors are used to registere unordered views of an object to construct its complete 3D model. Data is available for comparisons.
Integrating Local and Global Features for 3D Face Recognition
Local and global geometrical cues are integrated into in a single compact representation for 3D face recognition. The global cues provide geometrical coherence for the local cues resulting in better descriptiveness.
Research Grants
ARC (Discovery + ARF) Active Multispectral Computer Vision for Defense and Security Ajmal Mian $734,000 2011 - 2015
ARC (Linkage)* Automation of Species Recognition and Size Measurement of Fish from Underwater Stereo-video Imagery Euan Harvey, Mark Shortis, Ajmal Mian, Philip Culverhouse, Duane Edgington, Danelle Cline $436,000 2011 - 2014
ARC (Discovery) Person Identification using Multiple Non-invasive Iris and Face Biometrics in Video Robyn Owens, Ajmal Mian $390,000 2010 - 2012
ARC (Discovery + APD) Integration of Spatio-temporal Video Data for Realtime Smart Proactive Surveillance Ajmal Mian $311,298 2008 - 2010
DAAD Hyperspectral imaging for automatic handwriting analysis and segmentation Ajmal Mian, Andreas Dengel, Faisal Shafait $ 15,000 2014 - 2015
UWA (RCA) Biomass estimation of fish in the wild Ajmal Mian, Faisal Shafait, Bernard Ghanem $15,000 2014
UWA - FECM 3D Facial Morphometric Analysis for Objective Assessment of Craniofacial and Orthodontic Disorders Ajmal Mian $19,616 2012
UWA (RDA) Automatic Construction of 3D Deformable Fish Models for Biomass Measurement and Identification Ajmal Mian $28,087 2010
UWA (RDA) Fusion of Physiological and Behavioural Biometrics for Foolproof User Authentication Ajmal Mian $27,000 2007
*[Industry Partners: Australian Fisheries Management Authority, Dept. of Environment and Conservation, Dept. of Environment and Natural Resources, NSW Office of Environment and Heritage, Parks Victoria, SeaGIS Pty. Ltd] 

The Vice-Chancellor's Mid-Career Research Award 2014
West Australian Early Career Scientist of the Year  2012
UWA Outstanding Young Investigator 2011
Australian Research Fellowship (ARF), 2011
Australian Academy of Science ARC Award, 2009
Postdoc Award for Best Work Accepted for Publication, 2009
Australian Postdoctoral Fellowship (APD), 2008
Prize for Higher Degree by Research Achievements, 2007
Australasian Distinguished Doctoral Dissertation Award, 2006 (thesis)
Special Commendation for Higher Degree by Research Achievements, 2006
Postgraduate Research Travel Award, 2005

Dr. Yiqun Hu,             (2010-2011)
Dr. Arif Mahmood,     (Jan 2012 - Jan 2015)
Dr. Faisal Shafait,       (Jan 2013 - )

PhD Students (past)
Faisal R. Al-Osaimi (completed in 2010)     thesis title "Low-level fusion and deformation modeling for textured 3D face biometrics"
Yuk Lun Billy Li (completed in 2013, co-supervised with Wanquan Liu)     thesis title "Robust face recognition based on color and depth information"
Zohaib Khan (UWA, completed in 2014)     thesis title "Hyperspectral imaging and analysis for sparse reconstruction and recognition"
Ke Fan (thesis submitted in Dec 2014, co-supervised with Wanquan Liu)   
Mingliang Xue (thesis submitted Feb 2015, co-supervised with Wanquan Liu)    thesis title "Discriminant feature extraction and selection for person-independent facial expression recognition"

PhD Students (current)
Muhammad Uzair (since March 2012)        thesis title "Hyperspectral image set modelling"
Syed Zulqarnain Gilani (since May 2012)    thesis title "3D facial morphometric analysis for disease delineation"
Hossein Rahmani (since Nov 2012)            thesis title "View invariant human action recognition"
Naveed Akhtar (since March 2013)           thesis title "Bayesian sparse representation for hyperspectral image analysis"
Hasan Zaki (since Apr 2014)                     thesis title "RGB-D object and scene recognition"

Masters Student (current)
Paul McKay (since Jan 2015)                    thesis title "Facial expression recognition from 3D images"

Visiting Interns
Genping Zhao             (Harbin University of Technology China, 2014)
Zhipei Luo                  (USTC China, 2014)
Sheraz Ahmed            (DFKI Germany, 2014)
Muhammad I. Malik   (DFKI Germany, 2014)
Tiago Garcia              (Brazil Science Without Borders Student, 2013)

Conference Organization
Area chair for ACCV 2014, Singapore.
Program chair for DICTA 2012, Fremantle.

Peer Review
I have reviewed papers for IEEE Trans. on Pattern Anaylsis & Machine Intelligence (PAMI), Int. J. of Computer Vision (IJCV), CVPR, ECCV, BMVC, ACCV, IEEE Trans. on Circuits & Systems for Video Technology (TCSVT), IEEE Trans. on Image Processing,
IEEE Trans. on Information Forensics and Security, Image and Vision Computing (Elsevier), IEEE Trans. on Information Theory (IEEE IT), IEEE Trans. on Systems, Man & Cybernetics B (SMC), Signal Processing Journal (Elsevier), J. of Autism and Developmental Disorders,
Optics Express.