Selection of object detections using overlap map predictions
Md Sohel Rana, Aiden Nibali, Zhen He, “Selection of object detections using overlap map predictions.”, Neural Computing and Application, 2022.
CV
Education
- Ph.D. in Computer Science, La Trobe University, 2023
- Thesis title: Person Detection and Tracking in Dense Team Sports Scenes
- Description: During my PhD, I designed tracking algorithms for team sports such as hockey, netball, and volleyball. My work included developing models for team action recognition using player pose data. This experience allowed me to train and experiment with a variety of neural network models, including convolutional neural networks, recurrent neural networks, transformers, and more. This PhD was funded by Australian Institute of Sport.
- M.Sc. in Information and Communication Engineering Islamic University, Kushtia, Bangladesh, 2012
- Thesis Title: Improvement of Hijacking Problem for Multi-Target Tracking with Particle Filter
- Result: 3.57 out of 4.00 (1st Place-out of 40 students)
- M.Sc. in Information and Communication Engineering Islamic University, Kushtia, Bangladesh, 2011
- Project: Improved Eye Detection and Iris Center Tracking with Eyelashes Occlusion Correction
- Result: 3.52 out of 4.00 (1st Place-out of 40 students)
Work experience
- Machine Learning Engineer, Dynamic Crowd Measurement, AU, 2022-2023
- Developed a 2-dimensional lidar based crowd counting solution that can count people and their direction in live setup. These solutions were deployed in major events like Dutch grand prix, Melbourne Cup, Moto grand prix etc.
- Developed a real-time CCTV camera based people counting solution in jetson nano edge device.
- Developed a multi-modal sentiment prediction system that takes CCTV camera feed, social feed (i.e twitter stream) and weather data as input to a transformer backbone architecture and predicts crowd sentiments of an event. The Sydney Olympic Park Authority has implemented this system to monitor crowd sentiment during and post-events. project details
- Improved and deployed AI model for different crowd measurement task (e.g. crowd density, flow and sentiment) in different events.
- Trained, tested, validated computer vision-based CNN models and deployed those model using docker in AWS cloud/on-premise setup with maintaining MLOps best practices (e.g. model and data version management through S3, model serving, model monitoring and retraining etc).
- Designed python SDK based kafka consumer-producer along with AWS MKS for asynchronous data streaming.
- Performed causal inference analysis on crowd data for better understanding of crowd behaviour.
- Supervise junior engineers and review their PR requests.
- Data Scientist, Additive Assurance, AU, 2021-2022
- Evaluate the latest neural network model and engage in research aimed at enhancing the performance of defect segmentation in 3D printer layer images, with a specific focus on 3D printed bones.
- Conduct defect annotation of 3D printed images and train the model using annotated data
- Working with other team members to develop and deploy code.
- Lecturer, Dept. of Computer Science, Bangladesh University of Business & Technology, BD, 2016-2018
- Prepare lectures for courses and conduct classes (e.g. Neural Network, C++, Computer Graphics, Discrete Mathematics etc).
- Evaluate exam papers and supervise undergraduate student in their projects.
- Participate in academic and its relevant meetings.
Technical Skills
- Progragramming Languages and Libraries
- Python, C++, Matlab, PostgresSQL, NoSQL, OpenCV, Numpy, Pandas, CausalML, scikit-learn, APIs
- Model training, experiment and deployment tools
- PyTorch, Pytorch lightning, Tensorflow, CUDA for training, ONNX runtime model optimisation
- Llama 2, Langchan, CNN and RNN models, Transormers
- Cloud and Distributed computing
- AWS and GCP distributed computing
- Boto3 SDK for python, EKS, ECS, Sagemakers, Fargate, Lambda, EC2, VPC, S3, IAM, Secret Manager, MSK, Cloudwatch etc.
- Kubernetes (Deployment, Auto scaling, Load Balancing), Apache Kafka, Cassandra, Snowflakes
- Workflow and package management
- Docker, Kubeflow, MLflow, github action and bitbucket CI/CD pipeline, argflow, jenkins
- Documentation
- Latex, Inkscape for drawing, JIRA confuence page etc
- Extensive research and industry background of training Convolutional Neural Network (CNN) and Transformers across various tasks such as image classification. people tracking, eye tracking, segmentation, action recognition, multitask modeling, multimodal learning, text classification etc.
- Experienced in designing MLOps pipeline for model training, serving, monitoring, managing data drift, automated retraining.
- Proficient in diverse data science and machine learning tasks, including data analysis, causal inference, regression, k-mean clustering, Boosting algorithms, ensamble learning etc.
Other Skills:
- Working experience in result-driven and forward thinking work culture.
- Enthusistic about learning, sharing knowledge and applying new technologies.
- Collaboration and effective communication experience with diverse teams and a wide range of stakeholders both inside and outside the organization.
- Experience in leading teams, taking initiative, providing guidance, and upskilling both the team and myself.
- Accustomed to maintaining a daily schedule that includes attending stand-ups, brainstorming sessions, consulting with team members on technical decisions, performing code reviews, debugging, and handling pull requests.
- Enthusiastic about research, innovating new product ideas, and building proof of concepts.
Achievements:
- La Trobe University Graduate Research Scholarship for PhD (jointly funded by La trobe university and Australian Institute of Sport)
- Obtained scholarship for holding first position at B.Sc. and M.Sc. in order of merit.
- Completed Stanford online Machine Learning course (conducted by Andrew Ng) and obtained the certificate.
- Attended and held a position in National Collegiate Programming Contest [NCPC] 2011 at Dhaka University, Bangladesh.
Publications:
Automated repair of fragmented tracks with 1D CNNs
Md Sohel Rana, Aiden Nibali, Zhen He, Stuart Morgan, “Automated repair of fragmented tracks with 1D CNNs”, Image and Vision computing, 2020
Efficient multi-cue predictive tracking of multiple similar interacting objects
Rana, Md Sohel, et al. “Efficient multi-cue predictive tracking of multiple similar interacting objects.” 2018 Joint 7th International Conference on Informatics, Electronics & Vision (ICIEV).
Problem Analysis of Multiple Object Tracking System: A Critical Review
Md Zahidul Islam, Md Shariful Islam, Md Sohel Rana, “Problem Analysis of Multiple Object Tracking System: A Critical Review”, International Journal of Advanced Research in Computer and Communication Engineering, Vol. 4, Issue 11, November 2015.
Eye Detection And Iris Center Tracking With Eyelashes Occlusion Correction
Md Sohel Rana, Md Zahidul Islam, Md Shariful Islam, “Problem Analysis of Multiple Object Tracking System: A Critical Review”, International Journal of Advanced Research in Computer and Communication Engineering, Vol. 4, Issue 11, November 2015.
Talks
Person Detection and Tracking in Dense Team Sports Scenes
Thesis Defense Presentation at La Trobe University, Melbourne, Australia
(A workshop on deep learning and computer vision) Deep learning in a nutshell
Workshop at Islamic University, Kushtia, Banngladesh
Eye Detection And Iris Center Tracking With Eyelashes Occlusion Correction
Conference at Sudan University of Science and Technology, Khartoum, Sudan