Design PointNet Models, from First Principles to Deployment
Implement the architecture that started 3D deep learning, in PyTorch, layer by layer. 8 lessons for ML engineers who want to own the model, not import it.
See the architecture take shape
From an empty Python file to a trained PointNet classifying real 3D objects.
The year PointNet changed 3D ML. Every point-based architecture since builds on its core ideas.
Students trained worldwide across 80 countries.
Production experience distilled into structured, repeatable workflows.
You can import PointNet. Can you defend it?
Any ML engineer can pip install a 3D library and call a pretrained model. Then the interview question lands: “Why is max pooling the symmetric function here? What does the T-Net actually align? Why does your accuracy collapse on rotated scans?” Imports do not answer questions. Understanding does.
PointNet is the ideal architecture to truly own: small enough to implement in a few hundred lines, deep enough to contain the core ideas of all point-based learning, permutation invariance, shared MLPs, and learned alignment. Build it once properly, and every 3D paper you read afterward makes sense.
Architectures change every year. The reasoning behind them does not. When you have implemented the T-Net, debugged the pooling, and watched the loss curves yourself, you can evaluate any new 3D architecture in an afternoon. That judgment is what separates an ML engineer from a model consumer.
What you’ll build
A complete PointNet system, written by you, trained on real data.
PointNet, layer by layer
The full architecture in PyTorch: input and feature T-Nets, shared MLPs, and the max pooling that makes it order-invariant.
Data preparation pipeline
The unglamorous 80% done right. Point sampling, normalization, augmentation, and PyTorch datasets and loaders built for 3D.
Training + evaluation loop
Loss design, training dynamics, and the evaluation metrics that tell you whether the model generalizes or memorizes.
Clean engineering setup
A reproducible deep learning environment and a project structure you can reuse for every 3D model you build after this one.
Inference on real scans
Run your trained model on unseen 3D data, inspect the failure cases, and learn what they reveal about the architecture.
Architecture intuition
The mental model to read PointNet++, point transformers, and whatever comes next. The ideas transfer, the imports do not.
PointNet changed how we process 3D data in machine learning. If you are an ML engineer who wants to truly understand the architecture, not just import it, this is the course I wish I had when I started. We build everything, we break a few things on purpose, and you leave owning the model.
How this course works
Engineer-to-engineer, code-first, no hand-waving.
100% asynchronous
Access everything 24/7 on the LMS. Self-paced. No live sessions required.
From-scratch implementation
Every layer written and explained, not imported. You see why each design decision exists before you code it.
Real 3D datasets
Train and evaluate on real point cloud data, with the class imbalance and noise that benchmarks hide from you.
ML engineering patterns
Reproducible configs, clean dataset classes, and checkpointing habits that transfer to every model you ship.
Lifetime access
One payment, permanent access. Every future update included.
Gateway to 3D DL
PointNet is the foundation lesson of point-based deep learning. Everything in the 3D Deep Learning OS builds on what you learn here.
This is not a survey course. It assumes you can write Python, want to see tensors and shapes, and care about why the architecture works. If you want the broader segmentation method map instead, the Applied Semantic Segmentation course is the wider entry point.
The Curriculum
8 lessons. From an empty repo to a trained, evaluated PointNet.
Prerequisites
This course is for ML engineers and researchers who want to own point-based deep learning at the implementation level.
- Python (intermediate): comfortable with classes, NumPy, and reading library code
- ML basics: you know what a loss function and a training loop are. Prior PyTorch helps but is built up in the course
- Hardware: a GPU shortens training, but datasets are sized so a CPU still works. 16 GB RAM recommended
- Software: Python, PyTorch, NumPy, Open3D. All free and open-source
No prior 3D deep learning experience required. The architecture is built from first principles.
Start like you intend to ship. The starting folder, a reproducible deep learning environment, and the project structure the whole build lives in.
The heart of the course. T-Nets, shared MLPs, symmetric pooling, every component explained, then implemented in PyTorch.
Two full lessons on the stage that decides your accuracy. Sampling, normalization, augmentation, and PyTorch data pipelines for point clouds.
Close the loop. Model creation, training dynamics, inference on unseen scans, and the evaluation that tells you the truth about generalization.
Your instructor
Dr. Florent Poux
I’ve spent 12+ years in 3D geospatial: from field surveys with total stations to building AI systems for Fortune 500 companies. I published the O’Reilly book on 3D Data Science with Python. I’ve advised startups valued at over 15M EUR. I’ve held a professorship, taught at university, and led R&D for some of the largest organizations in the space.
I don’t teach syntax. I teach judgment. Every module is built around real decisions I’ve faced in production. Which neural renderer fits an industrial inspection job. How to architect a semantic pipeline that doesn’t choke on 500M points. When to use algorithmic methods and when to switch to deep learning.
What students say
Engineers, GIS professionals, and researchers from 80 countries.
“I came in fluent in Python but lost on 3D. After Module 2 I was reading PointNet++ papers like a normal human. The intuition Florent builds is the missing link.”
“We replaced a brittle classical pipeline with a sparse-conv network thanks to this course. Production accuracy jumped from 78 to 94 percent on the same data.”
“The training loop debugging chapter was worth the price by itself. I now know why my models stop learning instead of guessing at hyperparameters.”
“Implementing PointNet from scratch alongside Florent finally made the math feel like code. I caught a tensor-shape bug in our internal model the next day.”
“The PointNet++ to KPConv progression is taught the way I wish someone had taught me five years ago. Each architecture justifies the next, no hand-waving.”
“I needed to choose between sparse convolution backbones for a transit project. The architecture comparison module gave me a defensible recommendation in two days.”
Get lifetime access
One payment. The full architecture, the code, and every update.
Designing PointNet Models
Complete implementation curriculum + PyTorch source code + datasets + lifetime updates
- 8 lessons (8+ hours, 4 modules)i
- Complete PyTorch source code + datasets
- From-scratch architecture implementation
- Training, inference, and evaluation pipelines
- Lifetime access + all future updatesi
- 90-day results guaranteei
Zero-risk guarantee: If you don’t see real results within 90 days, I’ll refund you in full. No questions.
The complete ecosystem
3D AI Architect Program
The complete spatial AI curriculum, delivered in 3 tiers. Pick the depth that matches where you are — Foundations to get moving, Professional for the full OS stack, Ultimate for live access and priority support.
- 3D AI Acceleratori: 17 episodes in 6 acts
- 3D Course Libraryi: 24+ standalone courses
- All 4 OS courses (Professional & Ultimate tiers)
- Neurones 3D software access
- Monthly drop-in sessions with Dr. Poux (Ultimate)
- Spatial AI job and market intel
- Priority support + services access (Ultimate)
- 300+ hours of content
What you’re getting access to
Everything I’ve built over 12+ years, from land surveying in the field to advising 15M EUR startups, compressed into one curriculum you can start today. Delivered by the first QUALIOPI-certified 3D geospatial academy.
Every pipeline was battle-tested on Fortune 500 projects processing billions of points. You’re getting the real playbook, not theory.
Methods validated by peer-reviewed publications, the ISPRS scientific community, and 1,500+ academic citations. Not guesswork.
Built by someone who surveyed in the field, defended a PhD, advised funded startups, and shipped products to Fortune 500 clients.
I share more free content than most people put behind a paywall. That’s intentional. I want you to know exactly what you’re getting before you invest. This course is the concentrated, structured version of everything I know. No fluff. No filler. Just the production path.
Find the right path for you
From single courses to the complete ecosystem.
| Feature | Standalone Course | Designing PointNet Models | Course Library | 3D AI Architecti | Enterprise |
|---|---|---|---|---|---|
| Courses included | 1 topic | 4 modules | Full catalogi | 3 OS courses + Library (tiered) | Custom |
| Hours of content | 2-8h | 8+ hours | 150+ hours | 300+ hours (tiered) | Custom |
| Production source code | ✓ | ✓ | ✓ | ✓ | ✓ |
| Lifetime access | ✓ | ✓ | – | ✓ | ✓ |
| 3D AI Accelerator Tracki | – | – | – | ✓ | ✓ |
| Neurones 3D softwarei | – | – | – | ✓ | ✓ |
| Spatial AI job & market inteli | – | – | – | ✓ | ✓ |
| Monthly drop-in sessionsi | – | – | – | ✓ | ✓ |
| Priority support + services accessi | – | – | – | ✓ tiered | ✓ |
| Custom onboardingi | – | – | – | – | ✓ |
| Team licensing | – | – | – | – | ✓ |
| Price | €97 – €497 | €297 | €1,297 | Starts at €1,999 | On request |
Straight answers
Do I need prior PyTorch experience?
It helps but is not required. The course builds the PyTorch patterns it uses. You do need solid Python and a basic grasp of what training a model means.
What software do I need?
Python, PyTorch, NumPy, and Open3D. All free and open-source. No paid licenses required.
Do I need a GPU?
Training is faster with one, but the datasets are deliberately sized so a CPU works. 16 GB RAM recommended.
How long do I have access?
Lifetime. One payment, permanent access. Every future update is included.
What’s the refund policy?
90 days. Build the model, train it on your own data. If you don’t see results, email me for a full refund.
How is this different from 3D Deep Learning Foundations?
They share the PointNet core, but the framing differs. Foundations (35411) is the broader beginner on-ramp into 3D neural networks. This course is the tighter, engineer-focused implementation track. If you already work in ML and want the architecture itself, this is the direct path.
Is PointNet still relevant in 2026?
As a production model, sometimes. As the foundation of point-based learning, absolutely. PointNet++ and the transformer-based architectures all extend its core ideas. Understanding it from the inside is how you read every paper that came after.
Can I upgrade to an Operating System later?
Yes. When you are ready for the full production stack, contact me for credit toward the 3D Deep Learning OS or the Spatial OS.
Not sure if this course fits?
If you have specific questions about how the curriculum applies to your role, your team’s needs, or your technical background, I’m happy to help you figure it out before you commit.
Book a 15-min call