César Díaz Blanco

Hi there! I'm a machine learning master's student at the University of Tübingen. I am interested in using computer vision and its applications to media production from high quality and multi-scale 3D scene reconstruction to robust and dynamic human body representations.

I'm also a part-time working student in software engineering at DeepScenario. There I work on extending the detection and tracking pipeline to multiple and dynamic camera sensors. I've also worked on the release of inferred data by writing visualization and mesh manipulation scripts.

Email     Curriculum Vitae     Github     LinkedIn    

Dynamic Gaussian Splatting Viewer for Head Avatars

• Achieved on-the-fly curly hair editing with controllable amplitude and frequency
• Implemented a gaussian axes view with OpenGL using vertex and fragment shaders
• Enabled real-time frame loading for dynamic avatars and hair editing
Code  •   Demo  •   Slides  •   Report

Object detection with YOLO and ResNet-50

• Implemented YOLO's five-stage loss and fine tuned its regularization constants
• Reached 0.5175 mean average precision
Code  •   Demo

Neural network framework from scratch using NumPy

• Implemented backpropagation with support for sigmoid, softmax, ReLU, and linear layers
• Built in several optimizers including Adam, mini-batch SGD, and SGD
• Experimented by learning images with different positional encodings as seen in the demo
Code  •   Demo

NeRF and follow-up works

Report focusing on the volume rendering equation, its approximation in NeRF, and succesive works addresing dynamic scenes, efficient training, and surface extraction.
Written as part of a bonus exercise for the Virtual Humans WS23/24 course.
Report

Aliasing in Novel View Synthesis

Report and slides focusing on the shortcomings of NeRF and 3DGS when rendering images at different sampling rates than those in the training set. It gives a brief but detailed description of NeRF and 3DGS to understand where they fail, introduces Mip-NeRF and Mip-Splatting solutions, and highlights aliasing effects specific to differential rendering methods.
The report and slides were made as part of the Advanced Topics in Computer Graphics and Computer Vision Seminar WS24/25.
Report  •   Slides

TicGaToe

Reimagination of Connect4 on a circular board to play against your friends.
Get four of your checkers in a row forming a semicircle at any level or a line through the center.
Once a player has won, its checker will appear in the middle of the board.
Tap the center twice to play again.