Fraunhofer IIS — Production system
PyTorch DDP NCCL CUDA TensorRT DALI

End-to-end distributed training and inference pipeline for high-resolution surface inspection. Custom CUDA preprocessing, multi-node training with near-linear scaling, TensorRT INT8 deployment on edge devices.

32× A100 GPUs
1.8ms Inference
89% Scaling Eff.
99.2% Detection Rate
Compression Supports Spatial Deep Learning
TU Munich — PhD Research (IEEE JSTARS)
Image Compression CNNs VitisAI Remote Sensing

Aggressive image quantization and lossy compression as preprocessing for CNNs — reducing storage and training time with minimal accuracy loss.

19.9% Training Speedup
<1% Accuracy Loss
3 CNN Architectures
2 Satellite Datasets
ML Inference Benchmarking on FPGA
TU Munich — ESA-funded (w/ Airbus, OroraTech)
FPGA VitisAI TensorFlow INT8 Quantization Benchmarking

Benchmarked 11 CNN architectures on Xilinx ZCU102 FPGA for satellite onboard processing. INT8 quantization via VitisAI, power/throughput profiling.

11 Architectures
INT8 Quantization
ZCU102 Target FPGA
4 Publications
FPGA-Deployable CNN Model Library
TU Munich — ESA Project
TensorFlow VitisAI FPGA Model Zoo

Reimplemented 12 CNN architectures in TensorFlow with full VitisAI compatibility for FPGA deployment. ImageNet-pretrained weights included.

12 Models
77.9% Top Accuracy
ZCU102 Target HW
Trajectory Similarity via Compression
Bundeswehr University / TU Munich — Research
C++ Python pybind11 Kolmogorov Complexity Bloom Filters

Novel trajectory distance metric based on Normalized Compression Distance. Bloom filter embeddings compress trajectories to ≤1024 bits with >80% k-NN classification accuracy across real-world datasets.

≤1024 Bits/Trajectory
>80% Classification
O(n) Time Complexity
8.9% vs. Baselines
Satellite Change Detection via Kolmogorov Complexity
DLR / FH Salzburg — Master Thesis
Remote Sensing SAR Kolmogorov Complexity SVM

Parameter-free unsupervised change detection for multispectral and SAR satellite imagery using Kolmogorov complexity approximation. 70,000 km² processed with 36% runtime reduction.

70k km² Processed
36% Runtime Reduction
5 Publications