Moin Nabi

I am a Senior Research Scientist at SAP Machine Learning Research in Berlin.

Before that, I was a postdoctoral research fellow in Deep Relational Learning group at the University of Trento with Professor Nicu Sebe and a visiting researcher in the CS department at the University of Washington working with Ali Farhadi. I did my PhD at the Italian Institute of Technology where I was advised by Professor Vittorio Murino and also working closely with Professor Massimiliano Pontil from University College London.

I am so delighted to start computer vision research with Professor Mehrdad Shahshahani at IPM Vision Group. I did my masters in AI and my bachelors in software engineering in Iran.

Email  /  CV  /  Thesis  /  Google Scholar  /  LinkedIn


My research lies at the intersection of machine learning, computer vision, and natural language processing with an emphasis on learning Deep Neural Networks with minimal supervision and noisy/incomplete multi-modal data.

Internship Position: I like working with students. If you're a PhD student interested in a research internship working with me in Berlin, please send me an email with your CV and research interests.


[2017.12.01] We will present Differentially Private Federated Learning at Workshop on Machine Learning on the Phone and other Consumer Devices at NIPS 2017!

[2017.09.19] Our paper is awarded the Microsoft Best Student Paper in ICIP 2017.

[2017.05.15] I will join SAP Machine Learning Research in Berlin, as a Senior Research Scientist in Deep Learning.

[2017.04.01] Two papers are accepted in ACL 2017.

[2016.12.01] I will present Plug-and-Play Binary Quantization Layer at Workshop on Efficient Deep Learning at NIPS 2016!

[2016.11.11] The Motion Emotion Dataset (MED) is online!

[2016.09.25] Our work is finalist for the Best Paper Award in ICIP 2016.

Publications and Preprints

Self-Paced Deep Learning for Weakly Supervised Object Detection
E. Sangineto*, M. Nabi*, D. Culibrk and N. Sebe
arXiv:1605.07651v2, 2017  
PDF / code / bibtex

In this paper we propose a self-paced learning protocol for weakly-supervised object detection. The main idea is to iteratively select a subset of samples that are most likely correct, which are used for training. We show results on Pascal VOC and ImageNet, outperforming the previous state of the art on both datasets.

*Authors contributed equally


Plug-and-Play CNN for Crowd Analysis: An Application in Abnormal Event Detection
M. Ravanbakhsh, M. Nabi, Mousavi, E. Sangineto and N. Sebe
arXiv:1610.00307, 2016  
PDF / bibtex

We present a novel measure-based method which allows measuring the local abnormality in a video by combining semantic information (inherited from existing CNN models) with low-level Optical Flow. The proposed method is validated on challenging abnormality detection datasets and the showed the superiority of our method compared to the state-of-the-arts.


Emotion-Based Crowd Representation for Abnormality Detection
H.R. Rabiee, J. Haddadnia, H. Mousavi, M. Nabi, V. Murino and N. Sebe
arXiv:1607.07646, 2016  
PDF / dataset / bibtex

In this paper, we proposed a novel crowd dataset with both annotations of abnormal crowd behavior and crowd emotion. We also presented a method which exploits jointly the complimentary information of these two task, outperforming all baselines of both tasks significantly.


Sparse-coded Cross-domain Adaptation from the Visual to the Brain Domain
P. Ghaemmaghami, M. Nabi, Y. Yan and N. Sebe
IEEE International Conference on Pattern Recognition (ICPR), 2016   (Oral)
PDF / slies / bibtex

In this paper, an adaptation paradigm is employed in order to transfer knowledge from visual domain to brain domain. We experimentally show that such adaptation procedure leads to improved results for the object recognition task in the brain domain, outperforming significantly the results achieved by the brain features alone.


Novel Dataset for Fine-grained Abnormal Behavior Understanding in Crowd
H.R. Rabiee, J. Haddadnia, H. Mousavi, M. Kalantarzadeh, M. Nabi, V. Murino
IEEE Advanced Video and Signal-based Surveillance (AVSS), 2016  
PDF / poster / dataset / bibtex

This work presents a novel crowd dataset annotated by fine-grained abnormal behavior categoriy labels. We also evaluated two state-of-the-art methods on our dataset, showing that our dataset can be effectively used as a benchmark for the fine-grained abnormality detection problem.


CNN-aware Binary Map For General Image Segmentation
M. Ravanbakhsh, H. Mousavi, M. Nabi, M. Rastegari and C. Regazzoni
IEEE International Conference on Image Processing (ICIP), 2016  
Best Paper / Student Paper Award Finalist (7 / ~2000 submissions)
PDF / poster / slides / code / bibtex

In this paper we introduce a novel method for general semantic segmentation that can benefit from general semantics of Convolutional Neural Network (CNN). Experiments show that our segmentation algorithm outperform the state-of-the-art non-semantic segmentation methods by large margin.


Learning with Dataset Bias in Latent Subcategory Models
D. Stamos, S. Martelli, M. Nabi, A. McDonald, V. Murino, M. Pontil
IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2015  
PDF / abstract / bibtex

We present a multi-task learning framework which provides a means to borrow statistical strength from the datasets while reducing their inherent bias. In experiments we demonstrate that our method, when tested on PASCAL, LabelMe, Caltech101 and SUN in a leave-one-dataset-out fashion, achieves an average improvement of over 6.5% over state-of-the-arts.


Crowd Motion Monitoring Using Tracklet-based Commotion Measure
M. Nabi*, H. Mousavi*, H. Kiani, A. Perina and V. Murino
IEEE International Conference on Image Processing (ICIP), 2015  
PDF / poster / video / bibtex

We present a tracklet-based measure to capture the commotion of a crowd motion for the task of abnormality detection in crowd.

*Authors contributed equally


Abnormality Detection with Improved Histogram of Oriented Tracklets
H. Mousavi, M. Nabi, H. Kiani, A. Perina and V. Murino
International Conference on Image Analysis and Processing (ICIAP), 2015  
PDF / poster / bibtex

In this paper, we presented an efficient video descriptor for the task of abnormality detection in the crowded environments and achieved the-state-of-the-art results on the available datasets.


Webly-supervised Subcategoty-aware Discriminative Patch
M. Nabi, S. Divvala, A. Farhadi
Technical Report*, 2014  
PDF / slide / code

We study discovering a set of discriminative patches in subcategories of an object category, then train them in a webly-supervised fashion.

*This work was done while I was at UW.


Temporal Poselets for Collective Activity Detection and Recognition
M. Nabi, A. Del Bue, V. Murino
IEEE International Conference on Computer Vision Workshops, 2013   (Oral)
PDF / slide / video / code / bibtex

We present a mid-level motion representation based on actication patterns of Poselets detectors over time (Temporal Poselet), for detecting/recognizing the activity of a group of people in crowd environments.


Human Action Recognition in Still Images using Bag of Latent Poselets*
M. Nabi, M. Rahmati
9th European Conference on Visual Media Production (CVMP), 2012  
PDF / bibtex

We represent human body poses in a single images by extracting the Poselet activation vectors on it, and recognize human activities in still images using the proposed bag of latent Poselets.

*This work is based on my MS thesis at AUT.


Mid-level Representation for Visual Recognition
Moin Nabi
Ph.D. Dissertation , 2015  
PDF / slides / talk /

This thesis targets employing mid-level representations for different high-level visual recognition tasks, both in image and video understanding.


Stock trend prediction using Twin Gaussian Process regression
M. Mojaddady, M. Nabi, S. Khadivi
Technical Report, 2011  
PDF / bibtex


A Turorial on Digital Image Processing using MATLAB
M. Nabi
National Digital Image Processing Workshop, 2008  
PDF / code / data

Erdös = 3 (via two paths)

Thanks Jon!