I am a PhD candidate and a KUIS AI Fellow at Koç University, advised by Asst. Prof. Fatma Güney. My research focuses on 3D computer vision, with an emphasis on novel view synthesis, depth estimation, 3D reconstruction, and uncertainty estimation. I was also a visiting scholar at the University of Bologna, and have since been collaborating with Dr. Matteo Poggi and Dr. Fabio Tosi.
I received my Master's degree in Computer Science and Engineering from Koç University and my Bachelor's degree in Computer Engineering from Sharif University of Technology.
Email / CV / Google Scholar / GitHub / LinkedIn
FlowIt is a novel transformer-based global matching architecture for optical flow estimation, leveraging a hierarchical transformer design and an Optimal Transport-based flow initialization. FlowIt establish new state-of-the-art results on the Sintel benchmark, as well as zero-shot generalization performance across the Sintel, Spring, and LayeredFlow datasets.
WarpRF is a training-free framework for quantifying the uncertainty of radiance fields and can be applied to any radiance field implementation. WarpRF excels at both uncertainty quantification and downstream tasks, e.g., active view selection and active mapping, outperforming any existing method tailored to specific frameworks.
This paper addresses the poor geometric accuracy of 3D Gaussian Splatting (GS), which often produces inaccurate depth maps. It proposes a novel training strategy that incorporates depth priors from a stereo network by processing virtual stereo pairs rendered by the 3DGS itself, simultaneously improving the inferred 3D structure and the quality of novel view synthesis.
A motion-supervised approach for multi-object segmentation that requires only a single image at test time. By modeling optical flow as a low-dimensional function of depth and object motions, a network is trained that jointly estimates object masks and depth to explain the observed flow.
Self-supervised monocular depth estimation methods are typically scale-ambiguous. We show that the planar parallax formulation can resolve this limitation and train a self-supervised monocular depth estimation network producing metric outputs.
In this paper, we assume that there is an underlying process creating observations in a video and propose to factorize it into static and dynamic components. We model the static part based on the scene structure and the ego-motion of the vehicle, and the dynamic part based on the remaining motion of the dynamic objects. By learning separate distributions of changes in foreground and background, we can decompose the scene into static and dynamic parts and separately model the change in each.
MonoDepthSeg proposes to jointly estimate depth and segment moving objects from monocular video without using any ground-truth labels. It decomposes the scene into a fixed number of components where each component corresponds to a region on the image with its own motion, and estimates both the mask and the motion of each component.