Portfolio

Debojyoti Biswas

Machine Learning & Computer Vision Researcher

Results-driven ML researcher with 6+ years of experience building and deploying computer vision, robotics perception, autonomous systems, and multimodal AI solutions—bridging research prototypes to production with FastAPI, Docker, and AWS.

bishaldebojyoti@gmail.com +1 512-938-8033 LinkedIn CV (PDF)

Contact View Experience

State College, PA • Open to ML/Software roles

6+ yrs

ML/CV experience

CV • VLM

Perception & Understanding

MLOps

Docker • AWS • FastAPI

About

I build robust perception systems and scalable ML pipelines. My recent work spans multimodal 3D object detection for autonomous vehicles, domain adaptation for remote sensing, and vision–language models for long-duration video anomaly detection.

I enjoy shipping: packaging models into APIs, containerizing with Docker, and deploying on AWS (e.g., ECS/S3/SageMaker).

Focus: 3D perception, multimodal fusion, long-range reliability, explainable/robust ML.
Strength: research-to-production execution with clean, reproducible codebases.
Collaboration: cross-functional work with research, product, and engineering teams.

Skills

Coding & Frameworks

PythonC/C++PyTorchTensorFlow CUDAOpenCVHuggingFaceLAVIS

MLOps

DockerFastAPIAWS ECSAWS S3 SageMakerKafka

Data & Tools

NumPyPandasScikit-LearnMatplotlib MySQLSQLite

Experience

Aug 2025 — Present

Postdoctoral Researcher — Penn State University

3D Object Detection for Autonomous Vehicles in Adverse Weather

Built a multimodal 3D detection pipeline fusing sparse LiDAR + radar point clouds.
Introduced density/sparsity-aware feature extraction and neighborhood aggregation to improve context awareness.
Improved distant-object performance with an adaptive range-weighted loss; achieved +3.2% mAP on VoD.

Sep 2021 — Jul 2025

Doctoral Research Assistant — Texas State University

Remote sensing detection • Domain adaptation • VLM-based anomaly detection

Small Object Detection (CenterNet2)

Saliency-aware difficulty scoring + customized focal loss + heatmap proposals for dense small objects.
Achieved +1.8% / +2.3% mAP over SOTA on DOTA/DIOR.

Energy Efficient CNN / YOLOv5 Optimization

Compressed convolutions, backbone shrinkage, and scale-aware prediction for real-time UAV deployment.
Reduced GPU memory by 34% and power by 10W while maintaining accuracy (DIOR).

Domain-Adaptive Detection (GenAI intermediate domain + contrastive alignment)

Customized Detectron2 RCNN; improved +7.4% / +4.6% mAP on DOTA/NWPU-VHR10.

Cross-domain Detection (Debiased contrastive learning + pseudo-labeling)

Deployed with AWS (S3/SageMaker); improved mAP by +12.7% (VisDrone) and gains on DIOR/DOTA/UAVDT.

MMVAD / Long-duration Video Anomaly Detection

Designed vision–language anomaly detection with adaptive long/short segmentation and saliency-aware contrastive learning.
Deployed with Docker + AWS ECS + FastAPI; improved AUC on Drone-Anomaly/UIT-Drone/UCF-Crime/XD-Violence.

Jan 2019 — Jul 2021

Lecturer — Leading University

Taught 12+ undergraduate courses and labs.
Supervised 100+ student research projects; served on defense and evaluation committees.

Courses taught

Data Structures
Algorithms Analysis and Design
Object-Oriented Programming
C Programming
Computer Graphics
Computer Networking
Database Systems

Selected Publications

Small Object Detection from Satellite Imagery

Debojyoti Biswas, Jelena Tešić

Proposes difficulty-aware modeling and hard-example mining to improve dense small-object detection in satellite imagery.

DOI

BibTeX

@inproceedings{biswas2022small_object_satellite,
  title        = {Small Object Detection from Satellite Imagery},
  author       = {Biswas, Debojyoti and others},
  booktitle    = {International Conference on Computational Intelligence and Communication Networks (CICN)},
  year         = {2022},
  doi          = {10.1109/CICN56167.2022.10008383},
  url          = {https://doi.org/10.1109/CICN56167.2022.10008383}
}

NAS IEEE NAS (Networking, Architecture and Storage / NAS Conference) 2022

Energy Efficient CNN for Object Detection

Debojyoti Biswas, M. M. Mahabubur Rahman, Ziliang Zong, Jelena Tešić

Introduces efficiency-oriented CNN design choices for object detection, targeting reduced compute and improved deployment feasibility.
DOI
BibTeX
```
@inproceedings{biswas2022energy_efficient_cnn,
  title        = {Energy Efficient CNN for Object Detection},
  author       = {Biswas, Debojyoti and others},
  booktitle    = {IEEE NAS},
  year         = {2022},
  doi          = {10.1109/NAS55553.2022.9925528},
  url          = {https://doi.org/10.1109/NAS55553.2022.9925528}
}
```

Domain-Adaptive Small Object Detection in Satellite Imagery

Debojyoti Biswas, Jelena Tešić

Unsupervised domain adaptation for remote-sensing object detection using feature alignment and robust pseudo-labeling across dataset shifts.

DOI

BibTeX

@article{biswas2024domain_adaptive_small_object_detection,
  title        = {Domain-Adaptive Small Object Detection in Satellite Imagery},
  author       = {Biswas, Debojyoti and others},
  journal      = {IEEE Transactions on Geoscience and Remote Sensing},
  year         = {2024},
  doi          = {10.1109/TGRS.2024.3391621},
  url          = {https://doi.org/10.1109/TGRS.2024.3391621}
}

Debiased Contrastive Learning for Cross-Domain Object Detection

Debojyoti Biswas, Jelena Tešić

A debiased contrastive learning + pseudo-labeling framework to improve cross-domain detection robustness under geographic and sensor variations.

DOI

BibTeX

@article{biswas2024debiased_contrastive_cross_domain_detection,
  title        = {Debiased Contrastive Learning for Cross-Domain Object Detection},
  author       = {Biswas, Debojyoti and others},
  journal      = {IEEE JSTARS},
  year         = {2024},
  doi          = {10.1109/JSTARS.2024.3349541},
  url          = {https://doi.org/10.1109/JSTARS.2024.3349541}
}

Binarydnet53: a lightweight binarized CNN for monkeypox virus image classification

Debojyoti Biswas, Jelena Tešić

A novel lightweight binarized model, binarize the weights and biases of the DarkNet53, introduces large-margin feature learning and weighted loss calculation to enhance results.

DOI

BibTeX

@article{biswas2024binarydnet53,
  title={Binarydnet53: a lightweight binarized CNN for monkeypox virus image classification},
  author={Biswas, Debojyoti and Te{\v{s}}i{\'c}, Jelena},
  journal={Signal, Image and Video Processing},
  volume={18},
  number={10},
  pages={7107--7118},
  year={2024},
  publisher={Springer},
  doi= {10.1007/s11760-024-03379-8},
  url= {https://doi.org/10.1007/s11760-024-03379-8}
}

Springer SIVP Signal, Image and Video Processing 2024

Image deduplication using efficient visual indexing and retrieval: optimizing storage, time and energy for deep neural network training

M. M. Mahabubur Rahman, Debojyoti Biswas, Xiao Chen, Jelena Tešić

A novel framework for Visual Indexing and Retrieval-based image Deduplication (VIRD). VIRD effectively eliminates redundant data and maintains information quality in the training corpus through visual indexing and retrieval.
DOI
BibTeX
```
@article{rahman2024image,
  title={Image deduplication using efficient visual indexing and retrieval: optimizing storage, time and energy for deep neural network training},
  author={Rahman, MM Mahabubur and Biswas, Debojyoti and Chen, Xiao and Te{\v{s}}i{\'c}, Jelena},
  journal={Signal, Image and Video Processing},
  volume={18},
  number={12},
  pages={9495--9503},
  year={2024},
  publisher={Springer}
}
```

MMVAD: Multimodal Video Anomaly Detection in Drone and CCTV Footage

Debojyoti Biswas, Jelena Tešić

A vision–language anomaly detection approach leveraging captioning, attention-based fusion, and saliency-aware contrastive learning for surveillance videos.

DOI

BibTeX

@article{biswas2025mmvad,
  title        = {MMVAD: Multimodal Video Anomaly Detection in Drone and CCTV Footage},
  author       = {Biswas, Debojyoti and others},
  journal      = {Expert Systems with Applications},
  year         = {2025},
  doi          = {10.1016/j.eswa.2025.127857},
  url          = {https://doi.org/10.1016/j.eswa.2025.127857}
}

A Comprehensive Deep Learning Approach for Dermoscopic Image Enhancement

Abdullah Al Mazed, Md Faiyaj Ahmed Limon, Shahidul Haque Thouhid, Md Fazle Hasan Shiblee, Shubradeb Das, Md Shahid Iqbal, Debojyoti Biswas

EnhanceNet-U: a U-Net-based architecture for dermoscopic image enhancement which outperform existing CNN-based models

DOI

BibTeX

@article{al2025comprehensive,
  title={A comprehensive deep learning approach for dermoscopic image enhancement},
  author={Al Mazed, Abdullah and Limon, Md Faiyaj Ahmed and Thouhid, Shahidul Haque and Shiblee, Md Fazle Hasan and Das, Shubradeb and Iqbal, Md Shahid and Biswas, Debojyoti},
  journal={Franklin Open},
  pages={100405},
  year={2025},
  publisher={Elsevier}
}

Education

Ph.D. in Computer Science — Texas State University

Jul 2025 • CGPA 3.69/4.0

Thesis: Domain Adaptive Small Object Detection with Real-Time Response from High-Variance Remote Sensing Data

B.Eng. in Information & Communication Engineering — NSTU

Sep 2018 • CGPA 3.54/4.0

Contact

If you’re hiring for ML / VLM / CV perception roles, I’d love to connect.

Email LinkedIn Google Scholar CV (PDF)

bishaldebojyoti@gmail.com

Phone

+1 512-938-8033

Location

State College, PA