Berna Kabadayi

View My GitHub Profile

About me

Aloha! I am a PhD student at the Max Planck Institute for Intelligent Systems, supervised by Dr. Justus Thies and Prof. Gerard Pons-Moll. I am interested in 3D Computer Vision, Deep Geometry Learning, 3D Reconstruction & Tracking. Here is my resume.

Find me on GitHub or Twitter.


Master’s Thesis - Learning to Reconstruct 3D Objects by Part-based Template Deformation

We propose a learning-based method that reconstructs the 3D objects by part-based template deformation. Instead of retrieving the exact CAD model for a partial scan object, we first retrieve the most geometrically similar template by modeling a classification problem and then employing a bounding box regression to capture the object’s scale and translation. Lastly, our network learns part-based deformation of the retrieved template.


Metric Learning on Joint Embedding of 3D Scan and CAD Objects

We improve the scan-CAD retrieval baseline by 13% margin for the instance-level CAD model retrieval accuracy and by 18% margin for the Top-1 category-level CAD model retrieval accuracy by leveraging contrastive learning.

Research Report - Github


Face Reconstruction & Blend Shape Model

Given multiple RGB images of a face as an input, we use COLMAP to reconstruct the dense point cloud. Using landmarks, we fit the FLAME parametric face model to the reconstruction (Procrustes). Lastly, we deform the FLAME model then refine it further using as-rigid-as-possible deformation constraints. We implement a visualization tool that allows us to blend different expressions in real-time.

Final Presentation - Github - Blending Demo - Non-rigid Deformation Demo


Ray Tracer

I reimplemented the Peter Shirley’s Ray tracer in One Weekend series. Here is one of the outputs that I produced :)

Github - Ray tracer in One Weekend



E2EHDR is our senior project which supports capturing and displaying an HDR video stream in real time. Two cameras with different exposure settings are used for acquisition. Stereo matching is performed to align HDR frames in real time on the GPU. We have custom setup consists of LCD panel and projector for displaying.

We won the 4th place among 27 projects in the senior projects competition.

Find out more about E2EHDR