Lorenzo Porzi
Zurich, Zurich, Switzerland
Summary
🤩
Rockstar🎓
Top SchoolLorenzo Porzi is a research engineer at Meta Reality Labs with 11 years of experience specializing in photorealistic 3D reconstruction, semantic mapping and visual scene understanding. He built production-grade panoptic segmentation and memory-efficient training primitives at Mapillary—Seamseg and InPlace-ABN—work that influenced both academia and industry and is maintained as public code. His background spans PhD-level research in computer vision, postdoctoral mentorship, and practical deployments for mobile AR and large-scale street-level imagery. A hands-on ML engineer, he has implemented CUDA kernels and fixed subtle numerical issues in popular repositories, demonstrating deep low-level proficiency alongside algorithmic insight. Based in Zurich, he blends rigorous publications (ICCV/CVPR/PAMI) with production delivery for metaverse-scale 3D semantics. Beyond research, his varied early roles—from AR engine development to even acting—hint at a pragmatic, creative approach to problem solving.
11 years of coding experience
9 years of employment as a software developer
Doctor of Philosophy (Ph.D.), Computer Vision, Doctor of Philosophy (Ph.D.), Computer Vision at Università degli Studi di Perugia
Italian, English
Github Skills (13)
cuda10
gpu-programming10
thrust10
c-language10
activation-function10
batch-normalization10
cprogramming-language10
optimisation10
optimization10
deep-learning9
pytorch9
mask-rcnn7
faster-rcnn7
Programming languages (2)
Github contributions (5)
mapillary/inplace_abn
Nov 2017 - Nov 2021
In-Place Activated BatchNorm for Memory-Optimized Training of DNNs
Role in this project:
Back-end Developer & ML Engineer
Back-end Developer & ML EngineerContributions:16 releases, 70 commits, 1 PR in 4 years
Contributions summary:Lorenzo implemented CUDA-based activation functions, including leaky ReLU, ELU, and their backward passes, using the Thrust library for parallel processing. They addressed a potential NaN issue in the inplace_abn backward pass and fixed gradient calculation related to the weight. These changes involved modifying CUDA kernels and C++ code, indicating a focus on optimizing the performance and correctness of the in-place activated batch normalization module used for memory-optimized training of DNNs.
batchnorm
mapillary/seamseg
Jun 2019 - Nov 2021
Seamless Scene Segmentation
Contributions:38 commits, 1 PR, 24 pushes in 2 years 6 months
segmentation
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.