I build and evaluate computer vision and multimodal AI, from medical imaging to visual understanding.
Investigating the potential—and limits—of synthetic datasets for facial expression recognition.
Brain MRI segmentation under domain shift with GLIMS–MedNeXt. Conformal risk control is a separate, ongoing research direction.
Multimodal idiom understanding with IMMCAN. Studying how visual grounding and cross-attention connect image and language representations.
Our survey examines class-agnostic counting, from visual references to open-world text-guided approaches, and the challenge of generalizing to new objects.
Beyond the gallery / About
How can a model generalize responsibly when the real world does not match its training set?
I am a PhD student in Computer Engineering at Istanbul Technical University and a Graduate Research Assistant at SiMiT Lab, working under the supervision of Prof. Dr. Hazım Kemal Ekenel.
My research sits at the intersection of computer vision, deep learning, and trustworthy AI. I study how vision models behave beyond clean benchmarks: under domain shift, limited annotations, low-quality medical scans, privacy constraints, and imbalanced data.
Current directions include reliable brain-tumor MRI segmentation, conformal risk control, vision-language-guided facial expression generation, multimodal idiom understanding, and class-agnostic visual counting.
I enjoy exploring emerging ideas, expanding my knowledge, and applying what I learn to challenging real-world problems.
Research & education
Istanbul Technical University
Istanbul Technical University
SiMiT Lab, ITU
Istanbul Technical University
University of Tabriz
Technical expertise
Training and evaluation pipelines for image segmentation, facial expression recognition, and multimodal representation learning. Adapting pretrained models, comparing architectures, and evaluating generalization across datasets.
GPU-based workflows in Python and PyTorch, medical-image processing in MONAI, and experiments on Linux and HPC systems. Analyzing accuracy, segmentation quality, uncertainty, and failure cases.
Pretrained language and vision–language encoders, cross-attention and feature fusion, generative models for synthetic visual data, Hugging Face Transformers, and prompt engineering.
Additional areas of knowledge include large language models and retrieval-augmented generation (RAG), alongside graduate study in NLP, advanced NLP, and generative AI.
Three selected case studies
From the initial question to the experiment, findings, and limitations. Open a project to explore the details.
Adapting volumetric segmentation models from high-quality MRI to the imaging conditions represented in BraTS Sub-Saharan Africa.
MRI quality and data availability vary across imaging settings. A model trained on high-quality scans needs evaluation and adaptation before its results can be trusted on a different domain.
I developed and evaluated an ensemble of GLIMS and MedNeXt, using transfer learning, SSA fine-tuning, and fusion experiments to study segmentation performance under domain shift.
Data. BraTS GLI for pretraining → BraTS SSA for adaptation and evaluation; four MRI modalities.
The ensemble work was published in the MICCAI 2025 challenge proceedings in 2026. The paper reports improved accuracy and robustness on BraTS-SSA. See the publication for its full evaluation protocol.
Research evaluation on BraTS data; clinical deployment has not been established. Conformal risk-control experiments are a separate, ongoing research direction.
View the full evaluation ↗Building and comparing data-generation strategies for facial expression recognition when labels, class balance, and image-sharing constraints matter.
Expression datasets are imbalanced, and collecting or sharing face images raises practical privacy constraints. Can synthetic data improve transfer beyond the training dataset?
I investigated three dataset-construction strategies and evaluated their use in facial expression recognition: confidence-based pseudo-labeling, diffusion generation, and GAN expression editing.
Data. RAF-DB, FER2013, and AffectNet for evaluation; DigiFace, DCFace, EmoNet-Face BIG, and FFHQ as data sources.
IR50 with Mixed-SYN-C achieved 57.02% accuracy and 56.36% F1 on AffectNet, compared with 40.20% accuracy and 35.82% F1 for the RAF-DB-only baseline (Tables II and IV).
Results depend on the training mixture and target dataset. Synthetic-only training retains a domain gap; these scores are not a universal performance claim.
View the full evaluation ↗Ranking five image–caption candidates by how well they express the intended meaning of an idiom in context.
Literal image matching can miss an idiom’s intended meaning. The task also requires transfer across languages with limited supervision.
I co-developed a multimodal cross-attention framework integrating contextual language representations with visual and caption features for idiom understanding.
Data. AdMIRe 1.0 for supervised development; AdMIRe 2.0 for multilingual zero-shot evaluation. MAGPIE supports idiomaticity detection.
On the AdMIRe 2.0 ALL row, VTT-Cla-Base reached 0.350 accuracy and 0.727 NDCG, versus 0.288 and 0.704 for TT-Cla-Base (Table 3).
The text-only model performed better on the smaller AdMIRe 1.0 test. Caption augmentation had mixed effects. The linked repository currently presents the project overview; implementation availability should be checked there.
View the full evaluation ↗Selected publications / 2024–2026
Peer-reviewed work and open research across medical vision, facial analysis, multimodal reasoning, and visual counting.
Evaluates pseudo-labeling, diffusion synthesis, and GAN expression editing for facial expression recognition, including how synthetic and real data combine under cross-dataset evaluation.
@inproceedings{azmoudeh2026syntfer,
title = {On Applicability of Synthetic Datasets for Facial Expression Recognition},
author = {Ali Azmoudeh and Erdi Sarıtaş and Ömer Yıldırım and Hazım Kemal Ekenel},
year = {2026},
booktitle = {2026 IEEE 20th International Conference on Automatic Face and Gesture Recognition (FG)},
pages = {1--10},
publisher = {IEEE},
doi = {10.1109/FG67764.2026.11556946}
}Studies multilingual image ranking using frozen language and vision–language encoders with a trainable cross-attention module. Visual grounding improves zero-shot transfer, while text-only modeling performs better on the smaller in-domain test.
@inproceedings{bilen2026visaffect,
title = {VisAffect at MWE-2026 AdMIRe 2: IMMCAN Idiom Multimodal Cross-Attention Network},
author = {Barış Bilen and Ali Azmoudeh and Hazım Kemal Ekenel and Hatice Köse},
year = {2026},
booktitle = {Proceedings of the 22nd Workshop on Multiword Expressions (MWE 2026)},
pages = {149--153},
publisher = {Association for Computational Linguistics},
doi = {10.18653/v1/2026.mwe-1.19}
}Combines GLIMS and MedNeXt with pretraining, fine-tuning, and ensemble fusion for brain-tumor segmentation in lower-quality MRI.
@inproceedings{azmoudeh2026glimsmednext,
title = {GLIMS-MedNeXt: An Ensemble Framework for Brain MRI Segmentation in Sub-Saharan Africa},
author = {Ali Azmoudeh and İlkay Öksüz and Hazım Kemal Ekenel},
year = {2026},
booktitle = {Segmentation, Classification, and Synthesis for Brain Tumors and Traumatic Brain Injuries},
pages = {262--273},
volume = {16376},
publisher = {Springer Nature Switzerland},
doi = {10.1007/978-3-032-16365-3_24}
}Organizes 29 counting approaches into reference-based, reference-less, and open-world text-guided methods, comparing performance and generalization on FSC-147 and CARPK.
@article{ciampi2026counting,
title = {A Survey on Class-Agnostic Counting: Advancements from Reference-Based to Open-World Text-Guided Approaches},
author = {Luca Ciampi and Ali Azmoudeh and Elif Ecem Akbaba and Erdi Sarıtaş and Ziya Ata Yazıcı and Hazım Kemal Ekenel and Giuseppe Amato and Fabrizio Falchi},
year = {2026},
journal = {Computer Vision and Image Understanding},
pages = {104703},
volume = {267},
publisher = {Elsevier},
doi = {10.1016/j.cviu.2026.104703}
}Applies facial detection, landmark and feature extraction, and CNN–Transformer classification to examine expression responses during online shopping tasks.
@inproceedings{azmoudeh2024facialhci,
title = {Advanced Facial Expression Classification with CNN-Transformer Integration for Human-Computer Interaction},
author = {Ali Azmoudeh and Çiğdem Altın Gümüşsoy and Hazım Kemal Ekenel},
year = {2024},
booktitle = {2024 9th International Conference on Computer Science and Engineering (UBMK)},
pages = {800--805},
publisher = {IEEE},
doi = {10.1109/UBMK63289.2024.10773557}
}Research collaboration
For research discussions, collaborations, or questions about my published work, reach me through my ITU email.
azmoudeh22@itu.edu.tr