Bibesh Pyakurel

Research portfolio · Computer vision & multimodal evaluation

Applied AI for problems in the physical world.

I work on applied computer vision and multimodal model evaluation, with a secondary focus on benchmarking methodology for foundation models. I care about evaluation rigor: how task framing, prompt sensitivity, and metric selection change what we conclude about a model, and how conclusions survive the shift from controlled evaluation to deployment.

Read the papers Research statement

$research --focus→computer vision for infrastructure

Open to research collaborations and graduate study.

Portrait of Bibesh Pyakurel
Research · Software · Data
Affiliation
Research Lead, Dept. of Computer Science, UW–Green Bay
Advisor
Prof. Md Golam Sarwar Murshed
Preprints
2 on arXiv, 1 as first author
Looking for
Graduate study and research positions in computer vision and multimodal evaluation

01 / Statement

What I study, and why

My current work spans two projects. SLAPBench is the first structured benchmark for evaluating multimodal large language models on biometric image analysis. The pavement distress project builds instance-segmentation systems for real-world infrastructure assessment from field-collected imagery.

Across both, the question I keep returning to is whether an evaluation result means what it appears to mean. SLAPBench shows that a single change of prompt can collapse a verifier to near-total false acceptance while the same model separates classes almost perfectly under a different framing. The pavement work shows that pixel-level metrics and bounding-box metrics can tell different stories about the same predictions.

My goal is to develop evaluation frameworks and vision systems whose conclusions are reproducible and meaningful in low-resource, real-world contexts.

01

Computer vision for infrastructure

Instance segmentation and field-collected imagery for more precise pavement distress assessment.

02

Multimodal model evaluation

Benchmarks that reveal how prompting changes model behavior on demanding visual verification tasks.

03

Trustworthy biometric AI

Reliability, failure modes, and subgroup fairness in fingerprint verification with multimodal models.

02 / Papers

Publications

Two arXiv preprints. Each entry has a one-line takeaway, the headline numbers, the abstract, my role, and a BibTeX entry.

202601

Jul 17, 2026 · arXiv preprint · cs.CV, cs.AI · first author

SLAPBench: Benchmarking Multimodal Large Language Models for Four-Finger SLAP Fingerprint Verification

Bibesh Pyakurel · M. G. Sarwar Murshed

Prompting decides whether a multimodal model collapses on fingerprint verification; the model decides how well it can discriminate once the prompt lets it.

fingerprint pairs
7,832
multimodal models
5
prompting strategies
3
fairness attributes probed
3
Abstract

Four-finger SLAP images are the standard biometric capture format at US border entry points, yet no benchmark existed to test whether multimodal large language models can reason about them. SLAPBench builds 7,832 exhaustive pairs from NIST SD302b (176 mated, 7,656 non-mated) and evaluates InternVL3-8B, Qwen2.5-VL-7B, Qwen3-VL-8B, Gemma-3-12B, and Claude Opus 4.8 under zero-shot, task-description, and similarity-scoring prompts. Task-description prompting collapses every open-source model to near-total false acceptance; similarity scoring removes the collapse and exposes large differences in discrimination between models. The paper adds matched-resolution and demographic fairness probes over gender, race, and age, and treats a perfect AUC from one model as a diagnostic result because the dataset cannot rule out near-duplicate detection for mated pairs.

My role. First author. Designed the benchmark and protocol, ran all experiments, wrote the paper, and released the code and data.

arXivPDFDOICode & data
BibTeX
@misc{pyakurel2026slapbench,
  title         = {SLAPBench: Benchmarking Multimodal Large Language Models for Four-Finger SLAP Fingerprint Verification},
  author        = {Pyakurel, Bibesh and Murshed, M. G. Sarwar},
  year          = {2026},
  eprint        = {2607.15517},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2607.15517},
  url           = {https://arxiv.org/abs/2607.15517}
}
202602

May 25, 2026 · revised Jul 13, 2026 · arXiv preprint · cs.CV

Pixel-Level Pavement Distress Assessment Using Instance Segmentation

Logan Dewick · Bibesh Pyakurel · Kong Pheng Yang · Nazim Choudhury · M. G. Sarwar Murshed

Mask R-CNN with a ResNet-101 FPN backbone segments field-collected pavement distress well enough to recover crack area to within 0.006 points of ground truth.

F1, ResNet-101 FPN
87.04%
recall
90.04%
field images
1,874
pt. crack-area error
0.006
Abstract

The paper presents a Mask R-CNN instance-segmentation system for pixel-level pavement distress assessment, evaluated on the custom UWGB-StreetCrack dataset of field-collected images covering longitudinal, transverse, and alligator cracks plus potholes. Among the backbones compared, ResNet-101 FPN performed best with 84.23% precision, 90.04% recall, and 87.04% F1 under the paper's bounding-box matching protocol. Aggregate predicted crack-area fraction was 2.164% against 2.170% ground truth. A CSPDarknet53-based YOLO comparison reached 27.5% precision and 20.7% recall on the same data.

My role. Second author. Trained and compared the backbone variants, curated and annotated the dataset, and contributed to the write-up.

arXivPDFDOI
BibTeX
@misc{dewick2026pavement,
  title         = {Pixel-Level Pavement Distress Assessment Using Instance Segmentation},
  author        = {Dewick, Logan and Pyakurel, Bibesh and Yang, Kong Pheng and Choudhury, Nazim and Murshed, M. G. Sarwar},
  year          = {2026},
  eprint        = {2605.26095},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2605.26095},
  url           = {https://arxiv.org/abs/2605.26095}
}

03 / Research experience

Where the work happened

  1. Feb 2026 – Present

    Research Lead

    University of Wisconsin–Green Bay, Dept. of Computer Science · Advisor: Prof. Md Golam Sarwar Murshed

    • Designed SLAPBench as first author: 7,832 four-finger SLAP pairs from NIST SD302b (176 mated, 7,656 non-mated) and an evaluation protocol spanning five multimodal models and three prompting strategies.
    • Showed that prompting governs verification collapse while model capability governs discrimination; ran matched-resolution and demographic fairness probes; documented dataset limits that affect how perfect separation should be read.
    • Released the code and pair manifests publicly and led the pavement study through its arXiv revision.
  2. Sep 2024 – Dec 2024

    Research Assistant, Pavement Crack Detection

    WiSys Technology Foundation · Advisor: Prof. Md Golam Sarwar Murshed

    • Trained and compared five Mask R-CNN variants (ResNet-50/101 FPN, ResNet-101 C4/DC, ResNeXt-101 FPN) in Detectron2 on UWGB-StreetCrack (1,874 images, four distress classes). The ResNet-101 FPN model reached 84.23% precision, 90.04% recall, and 87.04% F1, five points over the ResNet-50 baseline.
    • Curated and annotated the dataset in Label Studio with polygon masks across longitudinal, transverse, and alligator cracks and potholes. Pixel-level predictions recovered a 2.164% crack-area fraction against 2.170% ground truth.
  3. 2024 – 2025

    Senior Research, B.S. Computer Science

    University of Wisconsin–Green Bay · Advisor: Prof. Md Golam Sarwar Murshed

    • WiSys-funded senior research on instance segmentation for pavement distress assessment, which became the arXiv preprint with Logan Dewick, Kong Pheng Yang, and Nazim Choudhury.

04 / Teaching

Teaching, mentorship, and recognition

Teaching & mentorship

  • Teaching Assistant · UW–Green BayJan 2024 – May 2024. Led weekly Python labs and office hours for 30+ students, graded with written feedback, and helped raise the lab pass rate from 20% to 45%.
  • Resident Assistant · UW–Green Bay HousingJan 2023 – Dec 2025. Advised a residential community on academic and personal challenges across three academic years.

Honors

  • Dean's List, Honors to Highest HonorsSix semesters, Fall 2022 – Fall 2025 · College of Science, Engineering and Technology, UW–Green Bay
  • Top Graduate, 2021Top 5 of 5,000+ graduates · Trinity International College, Nepal
  • WiSys research fundingSenior research on pavement distress assessment funded through the WiSys Technology Foundation

05 / Tools

Methods and tools

Modeling

  • PyTorch
  • Detectron2
  • Mask R-CNN
  • Hugging Face Transformers
  • scikit-learn
  • CUDA

Evaluation & data

  • Multimodal LLM prompting
  • ROC / AUC / EER analysis
  • Fairness probes
  • Label Studio
  • NumPy
  • Pandas
  • Matplotlib

Writing & reproducibility

  • LaTeX
  • Overleaf
  • Git
  • Linux / Bash
  • Docker
  • Python packaging

Education

  • B.S. Computer Science · University of Wisconsin–Green BaySep 2022 – Dec 2025

Coursework: Artificial Intelligence · Linear Algebra and Matrix Theory · Introductory Statistics · Data Structures · Discrete Mathematics · Database Design & Management · Cloud Computing · Computer Architecture · Advanced Software Design · Theory of Programming Languages

06 / Connect

Open to research collaborations and graduate study.

If you work on computer vision, multimodal evaluation, or benchmarking methodology and want to talk, email is the fastest way to reach me.

· GitHub · United States