Harald Scheidl
Computer Vision Engineer at Sportradar
Vienna, Austria
Summary
🤩
Rockstar🎓
Top SchoolHarald Scheidl is a Computer Vision Engineer based in Vienna with 10 years of experience building real-time vision systems that turn sports and industrial video into scalable, automated data pipelines. At Sportradar he leads a small engineering team and develops production-grade PyTorch and OpenCV solutions on AWS to extract live game data, while earlier roles saw him deliver embedded C++ and mobile vision components for product and spare-part identification. He is a hands-on ML engineer and open-source contributor, notable for implementing and optimizing CTC decoders and a TensorFlow-based handwritten text recognition system, and for improving the cppcheck static analyzer to catch subtle C/C++ issues. Comfortable across Python and C++, Harald combines low-latency optimization experience with practical dataset and tooling work—he even built synthetic-data renderers and a photo-capture app to bootstrap training sets.
10 years of coding experience
9 years of employment as a software developer
Vienna University of Technology
Stackoverflow
1,115reputation
49kreached
29answers
8questions
tensorflow
top-5%
Github Skills (39)
beam-search10
c-language10
static-analysis10
language-model10
python10
testing10
machine-learning10
recurrent-neural-networks10
c1110
decode10
c1710
mask-rcnn10
deep-learning10
tensorflow10
cplus10
Programming languages (6)
Github contributions (5)
githubharald/SimpleHTR
Jun 2018 - Dec 2020
Handwritten Text Recognition (HTR) system implemented with TensorFlow.
Role in this project:
ML Engineer
ML EngineerContributions:76 commits, 16 PRs, 70 pushes in 2 years 6 months
Contributions summary:Harald focused on implementing and improving the Handwritten Text Recognition (HTR) system. Their contributions included modifications to the command-line interface, integrating early stopping algorithms, incorporating data augmentation techniques during training, and optimizing the CNN kernel size for improved accuracy. Additionally, the user added features like validation mode, optional beam search, and character error rate calculations for evaluating the model.
text-recognitiondeep-learningrecurrent-neural-networksrecognitionhtr
githubharald/CTCDecoder
Oct 2017 - Jul 2021
Connectionist Temporal Classification (CTC) decoding algorithms: best path, beam search, lexicon search, prefix search, and token passing. Implemented in Python.
Role in this project:ML Engineer
ML EngineerContributions:2 reviews, 51 commits, 4 PRs in 3 years 10 months
Contributions summary:Harald primarily contributed to the implementation of various Connectionist Temporal Classification (CTC) decoding algorithms within the `ctcdecoder` repository. They focused on the creation and refinement of methods such as best path, beam search, and prefix search decoding. Their work involved modifications and additions to the `ctcDecoder.py`, `BeamSearch.py`, `TokenPassing.py`, and `PrefixSearch.py` files. Furthermore, the user addressed potential issues with repeated labels and improved the codebase for better performance.
passingpythondecodingbeamctc
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.