Michael Schwingshackl
@swini99.bsky.social
Researcher in 3D vision & robotics.
Synthetic data generation, geometry, and learning from imperfect reality.
Synthetic data generation, geometry, and learning from imperfect reality.
📄 Paper: arxiv.org/abs/2509.03262
🌐 Project: fafraob.github.io/pi3detr/
💻 Code: github.com/fafraob/pi3d...
🎮 Interactive demo: huggingface.co/spaces/fafra...
🌐 Project: fafraob.github.io/pi3detr/
💻 Code: github.com/fafraob/pi3d...
🎮 Interactive demo: huggingface.co/spaces/fafra...
PI3DETR: Parametric Instance Detection of 3D Point Cloud Edges With a Geometry-Aware 3DETR
We present PI3DETR, an end-to-end framework that directly predicts 3D parametric curve instances from raw point clouds, avoiding the intermediate representations and multi-stage processing common in p...
arxiv.org
January 29, 2026 at 1:35 PM
📄 Paper: arxiv.org/abs/2509.03262
🌐 Project: fafraob.github.io/pi3detr/
💻 Code: github.com/fafraob/pi3d...
🎮 Interactive demo: huggingface.co/spaces/fafra...
🌐 Project: fafraob.github.io/pi3detr/
💻 Code: github.com/fafraob/pi3d...
🎮 Interactive demo: huggingface.co/spaces/fafra...
Beyond benchmarks, PI3DETR is a strong foundation for:
🤖 robotic grasping
📐 3D measurement
🏗️ CAD reconstruction
🏭 digital manufacturing
🕶️ AR/VR
Works on real 3D Scans!
🤖 robotic grasping
📐 3D measurement
🏗️ CAD reconstruction
🏭 digital manufacturing
🕶️ AR/VR
Works on real 3D Scans!
January 29, 2026 at 1:35 PM
Beyond benchmarks, PI3DETR is a strong foundation for:
🤖 robotic grasping
📐 3D measurement
🏗️ CAD reconstruction
🏭 digital manufacturing
🕶️ AR/VR
Works on real 3D Scans!
🤖 robotic grasping
📐 3D measurement
🏗️ CAD reconstruction
🏭 digital manufacturing
🕶️ AR/VR
Works on real 3D Scans!
📊 Results:
• State-of-the-art performance on the ABC Dataset
• Robust across different curve parameterizations
• Simpler & more scalable than prior approaches
• State-of-the-art performance on the ABC Dataset
• Robust across different curve parameterizations
• Simpler & more scalable than prior approaches
January 29, 2026 at 1:35 PM
📊 Results:
• State-of-the-art performance on the ABC Dataset
• Robust across different curve parameterizations
• Simpler & more scalable than prior approaches
• State-of-the-art performance on the ABC Dataset
• Robust across different curve parameterizations
• Simpler & more scalable than prior approaches
PI3DETR extends 3DETR with a *geometry-aware matching & loss*.
This allows the model to:
✔️ handle differently parameterized curves
✔️ reason in continuous 3D geometry
✔️ stay fully end-to-end
This allows the model to:
✔️ handle differently parameterized curves
✔️ reason in continuous 3D geometry
✔️ stay fully end-to-end
January 29, 2026 at 1:35 PM
PI3DETR extends 3DETR with a *geometry-aware matching & loss*.
This allows the model to:
✔️ handle differently parameterized curves
✔️ reason in continuous 3D geometry
✔️ stay fully end-to-end
This allows the model to:
✔️ handle differently parameterized curves
✔️ reason in continuous 3D geometry
✔️ stay fully end-to-end
Most 3D curve / edge detection pipelines are:
❌ multi-stage
❌ heavily engineered
❌ tied to specific parameterizations
We asked: can we detect *parametric 3D curves directly* from point clouds in one shot?
❌ multi-stage
❌ heavily engineered
❌ tied to specific parameterizations
We asked: can we detect *parametric 3D curves directly* from point clouds in one shot?
January 29, 2026 at 1:35 PM
Most 3D curve / edge detection pipelines are:
❌ multi-stage
❌ heavily engineered
❌ tied to specific parameterizations
We asked: can we detect *parametric 3D curves directly* from point clouds in one shot?
❌ multi-stage
❌ heavily engineered
❌ tied to specific parameterizations
We asked: can we detect *parametric 3D curves directly* from point clouds in one shot?