Learned compression for 3D point clouds — M.Sc. Mechatronics @ FAU Erlangen-Nürnberg, thesis on neural point-cloud attribute codecs (sparse convolutions, multiscale entropy models, variable-rate conditioning).
🎯 Graduating 2026 · seeking roles in point-cloud / 3D compression & neural codecs · PyTorch MinkowskiEngine SLURM G-PCC / MPEG-PCC evaluation
My work has two halves, and I publish the tooling from both:
🔬 Evaluation you can defend — reproducing a state-of-the-art codec meant cross-calibrating two evaluation pipelines against the authors' released models until published numbers reproduced digit-for-digit. The discipline that came out of it lives in pcc-eval-toolkit: Bjontegaard metrics that refuse to produce numbers the data can't support.
⚙️ Training that survives real clusters — multi-week training fleets on a shared SLURM cluster with 24 h walltime caps, zero checkpoint loss. Patterns and scripts in slurm-resilient-training: relay chains, durable checkpoints, single-writer guards.
Earlier: learned prediction inside control loops — sEMG-driven tele-impedance (5 models × 3 horizons, MC-Dropout safety analysis).
📝 Research notes & longer write-ups: tianhao1314.github.io · 📫 Tianhao.Wu.Mechatronik@outlook.com
Thesis results and the full reproduction repo will be released with the thesis.