Building practical
computer vision
systems that ship.
Computer Vision Engineer specializing in real-time video analytics, multi-object tracking, segmentation, and perception system development — built end-to-end, evaluated rigorously.
Computer Vision Engineer with a B.Sc. in Computer Science from BRAC University, focused on building real-world perception systems across detection, segmentation, tracking, person re-identification, anomaly detection, and crowd analysis.
Each system is built with modular Python code, rigorous evaluation pipelines, and a strong focus on delivering working, production-quality CV systems — not just notebooks.
Languages
Python, Shell / Bash
CV / DL Frameworks
PyTorch, torchvision, timm, segmentation-models-pytorch, Ultralytics YOLOv8, boxmot
Computer Vision
OpenCV, detection, segmentation, tracking, ReID, anomaly detection, crowd density, optical flow, medical imaging
Model Architectures
EfficientNet, DenseNet, DeepLabV3+, Mask R-CNN, YOLOv8, CSRNet, ConvLSTM
Training Techniques
Transfer learning, fine-tuning, metric learning, class-weighted loss, cosine decay, ASPP, Grad-CAM
Tracking & ReID
StrongSORT, ByteTrack, OSNet, FAISS, cross-camera ReID, homography BEV, TrackEval
Evaluation
scikit-learn, AP/AR, mIoU, HOTA/MOTA/IDF1, CMC/mAP, MAE/RMSE, AUC-ROC, TensorBoard
Tools & Infra
Git, GitHub, Conda, Colab, Apple MPS, FAISS, ONNX export, real-time inference optimization, Albumentations
State-of-the-art multi-object tracking pipeline using YOLOv8m detection and StrongSORT with OSNet appearance embeddings, evaluated on MOT17 against a ByteTrack baseline. Includes track lifecycle management, velocity overlays, camera motion compensation, and JSON analytics.
End-to-end multi-camera surveillance pipeline with consistent person identities across 3 synchronized views. Cross-camera ReID uses cosine similarity and BEV projection on a floor map with trajectory animations and annotated output video.
Full person Re-ID system in pure PyTorch: ResNet-50 IBN-a, label-smoothing CrossEntropy, hard-mined triplet loss, and FAISS cosine search. Includes video demo pipeline with persistent global IDs and t-SNE visualization.
Dual-stream crowd anomaly system combining density and optical flow fed into ConvLSTM to detect panic, surges, and dispersal. Outputs annotated video overlays and anomaly score timelines for real-time surveillance analysis.
Unsupervised anomaly detection trained on normal surveillance frames only. Dual-channel autoencoder uses grayscale frames and optical flow to flag anomalies with heatmap overlays and scored clip extraction.
CSRNet for crowd density map regression on ShanghaiTech with adaptive Gaussian ground truth. Trained with CPU/MPS inference and includes a live heatmap demo with running count overlays.
Real-time tracking and counting pipeline with YOLOv8 and ByteTrack. Supports webcam and video streams, trajectory trails, movement heatmaps, virtual counters, and exports tracked results to CSV and JSON.
Traffic Sign Recognition on GTSRB
EfficientNet-B0 classifier across 43 classes with class-weighted training, webcam demo, and batch export support.
GitHubSemantic Segmentation on PASCAL VOC 2012
DeepLabV3+ segmentation system with ASPP, ResNet-50 backbone, and live inference for images and webcam overlay.
GitHubInstance Segmentation with Mask R-CNN on COCO
Fine-tuned Mask R-CNN with custom AP/AR evaluation, interactive webcam demo, and instance mask visualization.
GitHubMedical X-ray Classification on CheXpert
Multi-label DenseNet classifier with Grad-CAM explainability and uncertainty-aware training for chest pathology detection.
GitHubReal-Time Color Detection and Analysis App
Interactive OpenCV webcam app with HSV filtering, palette extraction, and live screenshot export.
GitHubBachelor of Science in Computer Science
BRAC University · Dhaka, Bangladesh
Class of 2025
Reference
Dr. Jannatun Noor Mukta
- Assistant Professor, CSE, BRAC University
- jannatun.noor@bracu.ac.bd
Let's build something that ships.
Open to remote roles in computer vision, video analytics, perception engineering, and applied deep learning. Available for full-time positions, contract work, and research collaboration.
Email Me