Ümit Mert Çağlar
PhD Candidate · Research and Teaching Assistant
Graduate School of Informatics, Middle East Technical University (METU), Ankara, Türkiye
Research Profile
My research focus is on computer vision and generative models, specifically addressing critical challenges in data scarcity, volume, and quality in remote sensing or medical imaging.
- Data Scarcity & Generative Augmentation: Utilizing Generative Adversarial Networks (GAN) and Denoising Diffusion Probabilistic Models (DDPM) to synthesize high-quality image data, evaluating their utility through extensive benchmarking and ablation studies.
- Geospatial & Earth Observation Benchmarks: Introduced two publicly available large-scale datasets: ARAS400k and BELDE.
- Efficient Deep Learning: Developed LALE (Lightweight-transformer Architecture for Land-cover Estimation), an efficient dense segmentation model.
- Data Quality Metrics: My recent research addresses data quality metrics crucial for tracking synthetic data generation fidelity and downstream model utility.
Selected Projects & Recent Work
ARAS400k
400k+ real & synthetic remote sensing images with segmentation maps and 2M+ vision-language grounded captions.
Read more → Dataset · ECCV 2026BELDE
1.08M curated image-mask pairs for RGB Earth-observation land-cover segmentation across Europe, with out-of-domain benchmarks.
Read more → Study · ECCV 2026Benchmarking Data-Quality Metrics
A human-perception study aligning automated quality metrics, human judgment, and downstream segmentation utility for synthetic EO data.
Read more → Model · ICMV 2026LALE
A lightweight hybrid convolution-transformer segmentation architecture reaching within 2.6 F1 of the best baseline at a fraction of the compute.
Read more → Study · ICMV 2025Colorectal Cancer Segmentation
Adaptive LLM-guided augmentation and multi-resolution ensemble models raising CRC tumor-grade segmentation F1 from 62.9 to 69.8.
Read more → Study · Earth Science Informatics 2025Single-Station Ground Motion
Exposing how deep learning models for epicentral distance estimation rely on auxiliary P/S phase timing rather than deep waveform features.
Read more →Education
PhD in Modelling and Simulation, Middle East Technical University · 2022 - Present Ongoing Research: Vision-Language Data Augmentation and Lightweight Segmentation Models for Scalable Remote Sensing Image Analysis.
MSc in Modelling and Simulation, Middle East Technical University · 2019 - 2022 Thesis: Improving Classification Performance of Endoscopic Images with Generative Data Augmentation.
BSc in Electrical and Electronics Engineering, Middle East Technical University · 2013 - 2018
Experience
Academic Experience (METU Graduate School of Informatics) - 2024 - Present
- Teaching Assistant, Transformers and Attention-Based Deep Networks
- Research Assistant, Computer Vision and Deep Learning Laboratory
- Researcher, Applied Intelligence Research Laboratory
Industrial Experience (TAI - Turkish Aerospace Industries) - 2019 - 2023
- Software Design Engineer, Data Science and Modelling
- Modeling and Simulation, Aeronautical-Avionics Simulations
- Technical Team Lead
STM - 2018
- RF and Simulation Systems Engineer
Awards and Accomplishments
- 2nd Place - ICIP Grand Challenge on Colorectal Cancer Tumor Grading and Segmentation (Out of 39 international teams) · 2025
- Best Poster Award - Graduate School of Informatics, 6th Open Research Day · 2025
- Special Mention - Graduate School of Informatics, 4th Open Research Day · 2023
- Best Poster Award - Graduate School of Informatics, 3rd Open Research Day · 2022
Scholarships
- TÜBİTAK 2211-A National Doctorate Scholarship · 2022 - 2027
- TEV (Turkish Education Foundation) Doctoral Scholarship · 2024 - 2027
Publications
(Full publication record available via ORCID and METU AVESIS)
- Grounding Synthetic Data Generation With Vision and Language Models, CVPR 2026 Synthetic Data for Computer Vision Workshop - Paper · Project Page
- LALE: Lightweight-Transformer Architecture for Land-Cover Estimation, ICMV 2026 - Paper · Project Page
- ARAS400k: A Large-Scale Remote Sensing Dataset Augmented with Synthetic Data for Segmentation and Captioning - Dataset · Project Page
- BELDE: Building a Large-scale Earth-observation Land-cover Dataset for Europe, ECCV 2026 Workshop - Paper · Project Page
- Benchmarking the Alignment of Data-Quality Metrics, Human Judgment and Land-Cover Segmentation Performance for Earth Observation, ECCV 2026 Workshop - Paper · Project Page
- Exploring Challenges in Deep Learning of Single-Station Ground Motion Records, Earth Science Informatics 2025 - Project Page
- Colorectal cancer segmentation with adaptive augmentation and multiresolution ensemble models, 18th International Conference on Machine Vision (ICMV 2025) - Paper · Project Page
- Progressive Disease Image Generation with Ordinal-Aware Diffusion Models, Diagnostics 2025 - Paper
- Colorectal Cancer Tumor Grade Segmentation in Digital Histopathology Images: from GIGA to Mini Challenge, IEEE International Conference on Image Processing Workshops (ICIPW) 2025 - Paper
- Class distance weighted cross entropy loss for classification of disease severity, Expert Systems with Applications 2025 - Paper
- Ulcerative Colitis Mayo Endoscopic Scoring Classification with Active Learning and Generative Data Augmentation, IEEE International Conference on Bioinformatics and Biomedicine (IEEE BIBM 2023) - Paper
- Improving classification performance of endoscopic images with generative data augmentation, METU 2022 - Thesis