Publications

A. Shafaei, J. J. Little, Mark Schmidt. Play and Learn: Using Video Games to Train Computer Vision Models. in BMVC, 2016.
[arXiv][PDF][MIT Technology Review][BMVC16][Abstract][Poster]
A. Shafaei, J. J. Little. Real-Time Human Motion Capture with Multiple Depth Cameras. in 13th Conference on Computer and Robot Vision, Victoria, Canada, 2016. (Oral Presentation)
[PDF][Project Page][Dataset][CNNs]
A. Shafaei, M.Sc. Thesis. Multiview depth-based pose estimation. in UBC Theses and Dissertations, 2015.
[PDF][cIRcle]
A. Gupta, A. Shafaei, J. J. Little and R. J. Woodham. Unlabelled 3D Motion Examples Improve Cross-View Action Recognition. in BMVC, 2014.
[PDF][Poster][Abstract][Project Page]

Curriculum Vitae

Use this link to download my curriculum vitae.

Undergrad Research

  1. A. Shafaei, “Design & Implementation of an Article Recommender System based on Probablistic Topic Models” Bachelor's Thesis in Software Engineering, Amirkabir University of Technology, Tehran, Iran, 2013.
  2. F. Einabadi, A. Shafaei, S. S. Ghidary, “3D Recognition of Hand-Object Activities in Home Environments” in proceedings of 4th RoboCup IranOpen International Symposium, Tehran, Iran, 2012.
    English: [PDF] [Abstract] [BibTeX]
    In this paper we present a novel system for recognition of hand activities where a person handles objects. For this end, cues from hand motion and object type are taken into account simultaneously. Hand motions are modeled with left-to-right Hidden Markov Models based on 3D motion features. Besides, objects are first detected and segmented using 3D depth data and then recognized by Scale Invariant Feature Transforms. Experimental results show the ability of the proposed system to recognize 5 different activities over 4 object types where noticeable similarities exists.
    Our main contributions in this work include introducing a new feature set for representing hand motions, showing right/left invariance for manipulating hand in the recognition task, and finally introducing a new hand-object scenario in home environments and recording its dataset.
    @inproceedings{EinabadiShafaeiGhidary12a,
      author = {Einabadi, F. and Shafaei, A. and Ghidary, S. S.},
      title = {3D Recognition of Hand-Object Activities in Home Environments},
      booktitle = {4th RoboCup IranOpen International Symposium},
      year = {2012}
      }
    }
  3. A. Shafaei, “Object Category Recognition Using a Probabilistic Context Model”, literature review of recent studies prepared as a presentation practice for Amirkabir University Technical Report Writings course, Tehran, Iran, 2012.
    Persian: [PDF][Poster PDF] [Abstract] [BibTeX] [Notes]
    This literature review was intended to be presented as a presentation practice for Technical Report Writings course at AUT, and does not carry original research nor maintains the typical literature review practice. Though I have tried my best to cite the original content where appropriate, but it may lack the sufficient citations.
    یکی از مسائل مطرح در حوزه‌ی بینایی کامپیوتر تشخیص دسته اشیاء یا طبقه‌بندی آنها است. هدف از طبقه‌بندی اشیاء شناسایی و یافتن نمونه‌های یک شیء در داخل تصاویر است.‬ ‫حل مسئله‌ی مطرح شده می‌تواند کاربرد زیادی در سیستم‌های هوشمند مانند روبات‌ها و یا سیستم‌های مبتنی بر تعامل با انسان داشته باشد.‬ ‫به طور کلی شناسایی اشیاء در عکس‌هایی که دارای همپوشانی، کیفیت پایین، نویز یا پس زمینه شلوغ هستند کار بسیار دشواری است. شناسایی زمانی دشوارتر می‌شود که تعداد زیاد اشیاء مختلف در صحنه حاضر شده باشند. مدل‌های بسیاری در طبقه‌بندی اشیاء از اطلاعات زمینه و ظاهری اشیاء برای بالا بردن دقت شناسایی استفاده می‌کنند. اطلاعات زمینه مبتنی بر نشانه‌های بصری می‌تواند دقت طبقه‌بندی را تا مقدار قابل توجهی افزایش دهد. اطلاعات صحنه مبتنی بر ارتباط میان اشیاء در آن، یا اطلاعات آماری کلی عکس می‌تواند ابهام ظاهری در شناسایی را نیز تا حد بالایی از میان ببرد. در این گزارش یک مدل کارآمد معرفی می‌کنیم که به کمک آن اطلاعات زمینه بیش از ۱۰۰ دسته‌ی اشیاء را با یک ساختار درختی-احتمالاتی پوشش داده‌ایم. مدل معرفی شده از ویژگی کلی تصویر، وابستگی‌های بین گروه اشیاء و همچنین تشخیص دهنده‌های محلی استفاده کرده و اطلاعات آن‌ها را در یک چهارچوب احتمالاتی قرار می‌دهد. نشان خواهیم داد که مدل زمینه ما قادر به افزایش دقت شناسایی اشیاء و همچنین قادر به تولید مفهوم منسجم از محیط است که به واسطه‌ی آن می‌توان جستجوی تصاویر را بر اساس گروه اشیاء و حالات صحنه انجام داد. سیستم ما همچنین قادر به تشخیص صحنه‌هایی است که تشخیص دهنده‌های محلی از آن عاجز هستند مانند تشخیص اشیاء خارج از صحنه یا جستجوی صحنه‌هایی با بیشترین مفهوم مشترک یا کمترین مفهوم مشترک. همچنین در این مقاله یک مجموعه‌داده‌ی جدید تحت عنوان ‏SUN معرفی خواهیم کرد که به منظور تشخیص زمینه و ارزیابی سیستم‌های مربوطه طراحی و آماده شده است. مجموعه‌ی داده شامل بیش از ۱۲۰۰۰ تصویر علامت خورده است که حاوی زمینه‌های بسیار زیاد و متفاوتی است (داخلی و خارجی) و همچنین بیش از ۲۰۰ دسته‌ی اشیاء و ۱۵۲۰۰۰ شیء علامت خورده را در برمی‌گیرد. آزمایشات اصلی این مقاله بر روی این مجموعه‌ی داده و همچنین مجموعه‌ی داده‌ی پاسکال صورت گرفته است.
    @misc{shafaei12a,
      author = {Shafaei, A.},
      title = {Object Category Recognition Using a Probabilistic Context Model},
      publisher = {Amirkabir University Technical Report Writings},
      year = {2012},
      note = {Literature review, in Persian}
    }
  4. A. Shafaei et al., “Sourena @Home 2011 Team Description Paper”, Technical Description Paper in RoboCup, Istanbul, Turkey, 2011.
    English: [PDF] [Abstract] [BibTeX]
    In this paper we introduce the "Sourena" robot for the 2011 RoboCup@Home competitions. We describe the hardware characteristics and software capabilities of Sourena which has made it reliable for different tasks in home environments.
    @misc{sourena11,
      author = {Ghaffarian, S. M. and Mahzarnia, H. and Vatankhah, H. and Einabadi, F. and Nejadgholi, M. and Shafaei, A. and Hosseini, S. M. and Shahsavari, H. and Kiomarsi, A. and Ghidary, S. S.},
      title = {Sourena @Home 2011 Team Description Paper},
      publisher = {RoboCup},
      year = {2011}
    }
  5. H. Mahzarnia, A. Shafaei and S. S. Ghidary “Finding objects bounding box and object recognition using stereo camera and robust local features for home robots” in proceedings of 6th Iranian Conference on Machine Vision and Image Processing, Isfahan, Iran, 2010.
    Persian: [PDF] [Abstract] [BibTeX]
    در سال‌های اخیر شناسایی اشیاء به یک مسئله مهم و اساسی در تحقیقات بینایی کامپیوتر بدل گشته است. در این پروژه حل مسئله شناسایی اشیاء برای روبات خانگی سورنا، مورد بررسی قرار گرفته است. شناسائی اشیاء توسط روبات در مقایسه با روشهای متداول مشکلات بیشتری را به همراه دارد. روبات باید به صورت خودکار اشیاء را در محیط یافته و از آنها نمونه‌برداری کند. نمونه‌ها در شرایط طبیعی و در محیط خانگی گرفته می‌شوند و اصولا اشیاء مورد بررسی کوچک هستند و در نتیجه با مسائلی مانند تغییرات در روشنایی محیط، نیاز به حذف زمینه از تصاویر، تغییرات در زاویه دید و تاری در تصاویر دوربین (بدلیل حرکت روبات و انسان در محیط) مواجه هستیم. برای حل مشکل نمونه برداری روشی ارائه گردیده است که ضمن حذف نمونه‌های با کیفیت پایین و تار، اشیاء را از زمینه آنها حذف کرده و کادر حاوی شیئ را محاسبه میکند؛ سپس با حرکت روبات نمونه‌گیری ادامه پیدا کرده و زاویه دیده شدن شیئ توسط روبات افزایش می‌یابد. همین طور با استفاده از ویژگی‌های مکانی مقاوم SIFT و ASIFT بر برخی دیگر از مسائل ذکر شده فائق آمده‌ایم. در نهایت با استفاده از روش تطبیق ویژگی، شناسایی اشیاء صورت گرفته است‪.‬
    @inproceedings{MahzarniaShafaeiGhidary12a,
      author = {Mahzarnia, H. and Shafaei, A. and Ghidary, S. S.},
      title = {Finding objects bounding box and object recognition using stereo camera and robust local features for home robots},
      booktitle = {6th Iranian Conference on Machine Vision and Image Processing},
      note = {In Persian},
      year = {2010}
    }
  6. A. Shafaei et al., “Sourena @Home 2010 Team Description Paper”, Technical Description Paper in RoboCup, Singapore, 2010.
    English: [PDF] [Abstract] [BibTeX]
    In this paper we introduce the "Sourena" robot for the 2010 RoboCup@Home competitions. We describe the hardware characteristics and software capabilities of Sourena which has made it reliable for different tasks in home environments.
    @misc{sourena10,
      author = {Mahzarnia, H. and Ghaffarian, S. M. and Rahat, M. and Vatankhah, H. and Hosseini, SM. and Einabadi, F. and Nejadgholi, M. and Shafaei, A. and Ghidary, S. S.},
      title = {Sourena @Home 2010 Team Description Paper},
      publisher = {RoboCup},
      year = {2010}
    }

Data Mining Research Laboratory (2011 - 2013)

Under Supervision Of
Prof. Khadivi

I joined this lab upon joining AUT Data Mining Team. It was founded by Dr. Khadivi in order of participation in International Data Mining Cups, and we were preparing for the 13th cup at the time. Research was mainly lead by the competition and was concerning sales forecast and dynamic price optimization, with real world data of 570 items during 42 days. More information about our achievements in that specific competition could be found here.

The methods we used included (but not limited to), various machine learning techniques, such as curve fitting, outlier detection, dynamical optimizations and neural network approaches as well as time series analysis.

Intelligent Data Analysis Research Laboratory (2012 Summer Internship)

Financial fraud detection (FFD) is an emerging topic of great importance, and this research was focused on this field of data mining. After a comprehensive literature review of the data mining techniques that have been applied to FFD, we decided to focus on creating a probabilistic platform to effectively employ all the available information at the research center. This research is still being refined and further developed. Significant results have been achieved by now, and the contributions in FFD will be published soon.

This research was supported by Behpardakht Mellat, a private local bank establishement.

Robotics Research Laboratory (2009 - 2011)

Under Supervision Of
Prof. Shiry

Research on object detection and recognition for home service robot Sourena which appeared in International RoboCup 2010 (Singapore), 2011 (Turkey) and papers appeared on 6th Iranian Conference on Machine Vision and Image Processing and 4th RoboCup IranOpen International Symposium.

As a young scholar, I started my journey during the freshman years here. During these two years I learned computer vision, machine learning and robotics in depth and practice. The most important problems that I was mainly involved with, concerned object recognition and object detection. At some point, I also studied object categorization techniques.

The main research was to prepare an effective solution for our home robot, Sourena in order to facilitate object recognition and object detection. Another research that I was present in, was the problem of activity recognition -recognition of human activity in home environment- in which we employed 3d-hand features in combination with graphical models. Our main contributions in this work included introducing a new feature set for representing hand motion, showing right/left invariance for manipulating hand in the recognition task, and finally introducing a new hand-object scenario in home environments and preparing its dataset.

By the end of the second year, as other masters team members graduated, I decided to further explore the world of machine learning in different areas.

More Pictures



Sample images of the dataset used for activity recognition.


I was responsible for object recognition service of our robot. The program that I developed mainly consisted of 3 stages. Preprocessing - Object Detection - Object Recognition. Since objects were supposed to be put on a surface (so robots could manipulate them) I employed a method to find surfaces, and extracted available objects afterwards. For object recognition I used the popular SIFT in combination with a color histogram method that I developed.