Projects
Things I've built, from research to tooling.
Large Language Models
-
LLM Explorer
Retrieval-augmented generation with LangChain, synthetic data from TinyLlama, and Hugging Face inference experiments.
Generative Models
-
Product Photography with Stable Diffusion and LoRA
Fine-tuning Stable Diffusion 1.5 with LoRA for product photos, filtering products with SAM and CLIP, and refining them with ControlNet inpainting.
-
Object Pose Editor with Stable Diffusion
Changing an object's pose in a photo: text-prompted segmentation, diffusion inpainting and novel-view synthesis.
-
Human Face Generator
A deep convolutional GAN trained from scratch on 200k CelebA faces, studying the effect of dropout.
Deep Learning for Computer Vision
-
PriorDepth
Joint monocular depth and camera-motion estimation from video, combining KeyPointNet and MonoDepth2.
-
Facial Keypoint Detector
A CNN in PyTorch that locates the eyes and mouth in face images.
-
Few-Shot Detection with Self-Supervised Backbones
Do self-supervised backbones (DINO, DetCO) help a Faster R-CNN detector when labelled data is scarce? A TUM Data Innovation Lab project with PreciTaste.
-
Semantic Segmentation of Pets
A U-Net trained from scratch with PyTorch Lightning on the Oxford-IIIT Pet dataset.
Machine Learning
-
Ensemble Methods
Soft-voting and stacking ensembles of random forests, extra trees and SVMs on MNIST.
Master Thesis
-
CLIPtheCenter
Master thesis: few-shot object detection by extending CenterNet with CLIP and DINO class codes.
Training Frameworks
-
Lightning Fast Trainer
YAML-configurable PyTorch Lightning recipes for training and evaluating image classifiers.
Traditional Computer Vision
-
Camera Tracking and Pose Estimation
Recovering object pose from image sequences with PnP, SIFT, RANSAC and reprojection-error optimisation.