|
Aloha! I'm a Ph.D candidate at Concordia advised by Prof. A. Ben Hamza. I am also an Applied ML Scientist at Décathlon where I help build AI/ML solutions that transform sport images and videos into actionable intelligence. Previously, I've also spent time at Ericsson as a Data Science Specialist. I obtained my MASc at Concordia where I focused on generative modeling (now Generative AI) for biomedical image understanding. Before that, I independently worked on computer vision research problems with Dr. Nabeel Mohammed, Dr. M. Sohel Rahman and Dr. Mahdy Rahman. 👀 During my bachelors at North South University (NSU) in Bangladesh, I founded and led the first IEEE Robotics and Automation Society (RAS) in Bangladesh when I dabbled in robotics, leading to IEEE BD RAS. 🤖 I like to read self-help books, workout, cook fancy meals and spend time in nature. I also like funny animal videos! 🐕 Email / LinkedIn / Google Scholar / GitHub / YouTube / Medium / Twitter |
|
|
My main focus is in the intersection of computer vision, machine learning and image processing with applications in visual perception and generation. I'm developing data-driven learning-based algorithms that enable computers to accurately and efficiently understand, model and recreate the visual world around us; all while reducing the need for intensive manual labeling required to build intelligent systems. 🧠 I have a strong first-author publication record in top-tier conferences in my field, like WACV and BMVC, high impact factor journals like Transactions in Medical Imaging and Computers in Biology and Medicine, as well as presented my research in prestigious conference workshops such as CVPR, ICML and MICCAI. For my Ph.D research, I recieved the Doctoral Graduate Fellowship in addition to the International Tuition Award of Excellence. In my MASc, I recieved the two-year MITACS Accelerate Fellowship. 🔬 I regularly contribute to open-source software and teaching. I have written feature code in YOLOv6, Kornia and a tutorial code in TensorFlow, Keras which also got featured in Henry AI Labs. I've also built utility and fun tools like Pynotify and Boss Detector. In my alma mater, NSU, I mentor (think Albus Dumbledore) undergrad baby scientists, specializing in machine learning which involves guiding with: coding/programming, algorithm implementations, generating research questions, designing experiments and paper writing. 😬 I support Slow Science. |
|
See also my Google Scholar profile for the most recent publications as well as the most-cited papers. |
|
RSUD20K is a new object detection dataset for road scene understanding, comprised of over 20K high-resolution images from the driving perspective on Bangladesh roads, and includes 130K bounding box annotations for 13 objects. We benchmark recent object detectors and explore Large Vision-Language models as image annotators. |
|
We propose a learning algorithm to explicitly focus on context from neighbouring regions around objects and learn a distribution of association across classes. Ideally to handle situations in-the-wild where only part of some object class is visible, but where us humans might readily use context to infer the classes presence. |
|
This paper introduces Cosine Similarity-based Image Filtering (CosSIF), a robust dataset filtering algorithm. We utilize CosSIF to create two filtering approaches: FBGT and FAGT. These methods rely on cosine similarity as the main metric for similarity calculation and aim to reduce the volume of GAN-generated synthetic images from the minority class that resemble similarity to images from the majority class. |
|
Masked Supervised Learning (MaskSup) is an effective single-stage learning paradigm that models both short- and long-range context, capturing the contextual relationships between pixels via random masking. Results show good segmentation performance, particularly at handling ambiguous regions and retaining better segmentation of minority classes with no added inference cost. |
|
FIFA is a a self-supervised conditional generative adversarial network based framework for virtual try-on. It consists of a Fabricator that aims to reconstruct the clothing image when provided with a masked clothing as input, and learns the overall structure of the clothing by filling in fabrics. A multi-scale structural constraint to enforce global context at multiple scales while warping the target clothing to better fit the pose and shape of the person. |
|
Automated identification of leukemia from microscopic image is crucial for early detection as the disease progresses rapidly. However, building these automated systems are challenging owing to fine-grained variability of lymphoid precursor cells and imbalanced data points. We benchmark several methods (input, loss amd GAN-based etc) designed to tackle class imbalance and find that loss-based methods outperform GAN-based and input-based methods for leukemia detection. |
|
Given a frame from a video sequence, we first segment single product item and hand followed by entropy masking to address the domain bias problem, and then a Vision Transformer (ViT) for multi-class classification. We also propose a custom metric that discards frames not having any product items, utilizing several image processing methods.Our method achieved 3rd place in the AI City Challenge, Track 4. |
|
Proposed a knowledge distillation approach for melanoma detection to reduce model complexitiy and enable easier deployment in edge devices. The proposed method requires relatively less time to detect melanoma. The method is 15 times smaller than EfficientNet-B0 and consistently achieves better performance in both melanoma and non-melanoma detection. |
|
Annotating medical image data is time consuming, costly and error prone, and the scarcity of labeled data limits the effectiveness of supervised learning. This thesis introduces generative modeling methods and architectures to address the scarcity of labels as well as class imbalance which results in a bias towards to over-represented class for tasks such as classification, binary and multi-class semantic segmentation. |
|
Sharp U-Net achieves improved performance on six medical image segmentation datasets in both binary and multi-class segmentation tasks while adding no extra learnable parameters. The idea is instead of applying a plain skip connection, a depthwise convolution of the encoder feature map with a sharpening kernel filter is done prior to merging the encoder and decoder features. |
|
An efficient semi-supervised learning method is proposed for binary and multi-class image classification. The method requires 6x less compute time and 5x less memory compared to prior arts. We also show that our method boosts robustness of visual classification models, even without specifically optimizing for adversarial robustness. |
|
We pretrain a model on videos for human activity recognition which leads to better representations for the downstream tuberculosis type classification task, especially for under-represented class samples. Our method achieved 2nd place in the ImageCLEF 2021 Tuberculosis Type Classification Challenge. |
|
Automating the tasks of detecting, segmenting, and classifying cell nuclei can free up the pathologists’ time for higher value tasks and reduce errors due to fatigue and subjectivity. This paper summarizes and publicly releases the challenge dataset, and compile key findings of the methods developed by various participants. |
|
Propose synthetic image data by leveraging class conditioning and adversarial training that achieve results comparable to training with only real data when using a test set of real images. Also combine synthetic data with different sizes of real datasets for additional performance gains. |
|
This work proposes a blood cell segmentation dataset consisting of multiple cell types. Additionally, all cell types do not have equal instances, which encourages researchers to develop algorithms for learning from imbalanced classes in a few shot learning paradigm. We also provide both learning and non-learning based methods as baselines. |
|
Transformed a high variation malaria localization dataset into a malaria classification dataset. Several classification algorithm benchmarks are provided for the task of classifying presence of malaria from microscopic images of isolated red blood cells. |
|
Showed that analyzing 3D medical images in a per slice basis is a sub-optimal approach, that can be improved by 3D context. Ranked 5-th in ImageCLEF 2019. |
|
Improved classification performance by synthesizing under-represented class samples from the over-represented ones. Synthetic samples are used as additional training data to reduce class imbalance. |
|
SincNet models based on joint angular margin loss not only consistently outperformed current prior models, but also generalizes well on unseen and diverse tasks such as Bengali speaker recognition. |
|
Benchmarked several classification algorithms for the task of detecting malaria from microscopic images of red blood cells. Transfer learning approach worked best in our study. |
|
A 3D CNN with a slice selection method employed in the task of chest CT image analysis for predicting tuberculosis (TB). Our method achieved 10-th place in the ImageCLEF 2019 Tuberculosis SVR - Severity scoring. |
|
An accuracy of 97.09% was achieved on the NumtaDB Bengali handwritten digit dataset, which was obtained by freezing intermediate layers. |
|
Implementation of an automated multifunctional attendance system which uses rfid, fingerprint, and real time facial recognition. |
|
A wireless sensor network with a real time web application for monitoring multiple ships to prevent catastrophic events due to overloading. |
|
A low-cost autonomous mobile security robot based on a multisensor system for the purpose of sending alarms remotely. |
|
|
|
Res-M3D-VTON is a pipeline for monocular to 3D virtual try-on (VTON) for fashion clothing which uses residual learning to synthesize correct clothing parts, preserve logo of clothing and reduce artifacts to finally output better textured 3D try-on meshes. |
|
AI Model Code / REST API Code
This is a demonstration of a full stack deep learning project from training a model to deploying it, using a REST API endpoint as well a separate end-user prototype web application. The model is built for predicting the presence of melanoma from dermoscopic skin lesions using neural networks. |
|
Collaborators: Aimon Rahman fastMRI Image Reconstruction Challenge, 2019 We use deep encoder-decoder architectures to reconstruct a high resolution knee MRI image given a low resolution MRI image. |
|
Collaborators: Tajwar Abrar Aleef, Aimon Rahman and Labib Chowdhury MICCAI TN-SCUI Challenge, 2020 In this work, several state-of-the-art image segmentation techniques were explored for optimizing Thyroid Nodule segmentation from Ultrasound images. This inlcuded supervised learning, transfer learning, and generative adversarial learning. |
|
Product Counting and Recognition for Retail Checkout, AI City Challenge, CVPR Workshop, 2022 (3rd Place) Tuberculosis Type Classification, ImageCLEF, 2021 (2nd Place) Nuclei Segmentation and Classification, MoNuSAC, 2020 (11th Place) Tuberculosis Prediction, ImageCLEF, 2019 (5th Place) Bengali Handwritten Digit Recognition, Bengali.AI, 2019 (6th Place) |
|
|
|
|
RSUD20K is a new object detection dataset for road scene understanding, comprised of over 20K high-resolution images from the driving perspective on Bangladesh roads, and includes 130K bounding box annotations for 13 objects.. |
|
|
The dataset consists of 21,295 synthetic COVID-19 chest X-ray images generated using algorithm. Dataset is publicly available here. The primary use of this dataset is to be used as additional data for training machine learning models. |
|
|
 |
Lab Demonstrator, COMP 6771: Image Processing, Winter 2022
Lab Demonstrator, COMP 333: Intro to Data Analytics, Fall 2021 Lab Demonstrator, COMP 6771: Image Processing, Winter 2021 |
|
Build and Deploy Custom Docker Images for Object Recognition, Towards AI , 2023Deep Learning in Computer Vision with PyTorch , 2023Intro to Deep Learning with NumPy, NSU , 2022Building ML models with TensorFlow, Ericsson Canada , 2021How to get started with building Computer Vision systems, NSU , 20213D image classification from CT scans, Keras, TensorFlow , 2020Programming with Python, NSU , 2019Intro to Deep Learning for Image Classification using Python, NSU , 2019Basics of Image Processing and Computer Vision, NSU , 2018Intro to Robotics (ROBO101), a semester-long series of workshops, NSU, 2018 |