Zeynep Sonat Baltacı

I am a PhD student in the Imagine lab at ENPC, under the supervision of Mathieu Aubry. My work specializes in developing interpretable vision models for clustering and object discovery, with an application to digital humanities 📜.

Prior to my PhD, I received my M.Sc. degree from METU, with a focus on long-tailed image recognition under the supervision of Emre Akbaş and Sinan Kalkan.

In my free time, I enjoy sightseeing 📷, dancing 💃🏼, and musical theatres 🎭.

           

News 🗞️

Publications 📝

⭐ indicates selected publications.

Diagrams Image
Text Region Detection in Historical Astronomical Diagrams ⭐
Z. S. Baltacı*, R. Baena*, F. Meng, S. Norindr, F. Somer, M. Husson, M. Aubry
Webpage ArXiv GitHub Hugging Face Poster - ICDAR 2026

We introduce a large-scale, diverse, open-access dataset of 948 historical astronomical diagrams containing 10,940 oriented polygonal text regions spanning ten centuries and seven linguistic traditions.

DS Image
Deep Sprite-based Image Models: An Analysis ⭐
Z. S. Baltacı, R. Loiseau, M. Aubry
Webpage ArXiv GitHub - TMLR 2026

We perform an extensive analysis of sprite-based image models on clustering benchmarks and propose a deep decomposition method that scales linearly with the number of objects.

PoM Image
PoM: A Linear-Time Replacement for Attention with the Polynomial Mixer
D. Picard, N. Dufour, L. Degeorge, A. Ghosh, D. Allegro, T. Ravaud, Y. Perron, C. Sautier, Z. S. Baltacı, F. Meng, S. Kalleli, M. López-Rauhut, T. Loiseau, S. Albouy, R. Baena, E. Vincent, L. Landrieu
ArXiv GitHub - CVPR Findings 2026

This paper introduces the Polynomial Mixer (PoM), a novel token mixing mechanism with linear complexity that serves as a drop-in replacement for self-attention.

ID Image
Intrinsic Dimensionality as a Model-Free Measure of Class Imbalance
Ç. Eser, Z. S. Baltacı, E. Akbaş, S. Kalkan
ArXiv GitHub - Neurocomputing 2026

We propose an alternative perspective on imbalance in long-tailed datasets, focusing on the intrinsic dimensionalities of classes in image space rather than their cardinalities.

ROIi Image
Historical Printed Ornaments: Dataset and Tasks ⭐
S. K. Chaki*, Z. S. Baltacı*, E. Vincent, R. Emonet, F. Vial-Bonacci, C. Bahier-Porte, M. Aubry, T. Fournel
Webpage ArXiv GitHub Hugging Face Poster - ICDAR 2024

We propose a historical printed ornament dataset associated with three different tasks: clustering, element discovery and unsupervised change localization.

IAMAHA Image
Computer Vision and Historical Scientific Illustrations
F. Aouinti, Z. S. Baltacı, M. Aubry, A. Guilbaud, S. Lazaris
Webpage HAL GitHub - IAMAHA 2023

We present a semi-automatic interactive pipeline for scientific illustration extraction and introduce a new dataset of scientific illustrations.

CU Image
Class Uncertainty: A Measure to Mitigate Class Imbalance
Z. S. Baltacı, K. Öksüz, S. Kuzucu, K. Tezören, B. K. Konar, A. Özkan, E. Akbaş, S. Kalkan
ArXiv

To measure class imbalance, we propose "Class Uncertainty" as the average predictive uncertainty of the training examples, and we show that this novel measure captures the differences across classes better than cardinality.

maIoU Image
Mask-aware IoU for Anchor Assignment in Real-time Instance Segmentation
K. Öksüz, B. C. Çam, F. Kahraman, Z. S. Baltacı, S. Kalkan, E. Akbaş
ArXiv GitHub - BMVC 2021

We propose Mask-aware IoU, an IoU variant for better anchor assignment to supervise instance segmentation methods.

Generalized Mask-aware IoU for Anchor Assignment in Real-time Instance Segmentation
K. Öksüz, B. C. Çam, F. Kahraman, Z. S. Baltacı, S. Kalkan, E. Akbaş
ArXiv

Talks 💬

Academic activities 👩🏼‍🎓

Contact