[Seminar] Physically Interactable 3D Indoor Scene Synthesis

preview_player
Показать описание
Speaker: SeungWon Seo
Рекомендации по теме
[Seminar] Physically Interactable 0:17:50

[Seminar] Physically Interactable 3D Indoor Scene Synthesis

Interactive Physically-Based Sound 0:05:20

Interactive Physically-Based Sound Design of 3D Model using Material Optimization

[AAAI 2023] Pose-guided 0:02:54

[AAAI 2023] Pose-guided 3D Human Generation in Indoor Scene

3D Semantic Parsing 0:11:43

3D Semantic Parsing of Large-Scale Indoor Spaces

3D indoor map 0:00:57

3D indoor map library curation interface and examples

[MERL Seminar Series 1:06:04

[MERL Seminar Series 2021] Learning to See by Moving: Self-supervising 3D scene representations ...

Lumion Phasing Animation 0:00:13

Lumion Phasing Animation - Realistic Isometric Interior 3d Render #Shorts

High security locks 0:00:25

High security locks easily defeated w/ this method. #lockpicking #homeowner #tools #city #security

NCI Seminar: 3D 0:58:38

NCI Seminar: 3D Imaging of Fruit Fly Neurons and Muscles

Fake 3D animation 0:00:31

Fake 3D animation in figma. Get the figma file from our figma community, link in bio.

GridDrones: A Self-Levitating 0:17:55

GridDrones: A Self-Levitating Physical Voxel Lattice for Interactive 3D Surface Deformations

WAML 2021 | 0:49:49

WAML 2021 | Indoor Space Modeling with 3D Web GIS for Library Asset Management, Discovery, and More

Interactive 3D scientific 0:29:26

Interactive 3D scientific data visualization and mesh analysis with PyVista - Jan Margeta

How to Answer 0:00:27

How to Answer Any Question on a Test

3D hologram fan 0:00:23

3D hologram fan portrait solution. Who wanna date this holographic sexy lady #3dhologramfan

Revision 2016 - 0:55:38

Revision 2016 - Seminar - BeRo & urs: The design of the KRAFT 3D rigid body physics engine

ProtoHole: Prototyping Interactive 0:00:30

ProtoHole: Prototyping Interactive 3D Printed Objects Using Holes and Acoustic Sensing

Workshop: 3D Content 2:42:59

Workshop: 3D Content Creation for Simulated Training in eXtended Reality (TrainingXR) | IEEE VR 2022

QtWS17 - Interacting 0:49:45

QtWS17 - Interacting with 3D content, Mike Krus, KDAB

How to Answer 0:00:31

How to Answer Any Question on a Test

Macrophage eating Cancer 0:00:11

Macrophage eating Cancer Cell 3D Animation | Phagocytosis | Immunology @biologyexams4u

Workshop: 3D Reconstruction, 3:13:53

Workshop: 3D Reconstruction, Digital Twinning, and Simulation for Virtual Experiences (ReDigiTS)

Scratch Conference 2024 0:54:42

Scratch Conference 2024 Workshop - Unplugged! Physical 3D Scratch Blocks: Lessons from Ghana

What is depression? 0:04:29

What is depression? - Helen M. Farrell

Комментарии
Автор

Thank you for the great presentation!
I found your paper really interesting, so I ended up with quite a few questions—hope that’s okay:

Q1. In Slide 9, you mentioned that the conditioned diffusion model takes time embedding and floor plan embedding as conditions. How are these embeddings obtained, and what forms do they take? I'm particularly curious about the floor plan embedding.

Q2. On Slide 9, you explained that the conditions are added to all layers. Does this mean they are added through simple addition calculations, as opposed to the attention mechanism used in stable diffusion, where conditions influence the attention computations for each layer's output?

Q3. Could you provide a more intuitive explanation of the modifications to the denoising process for constraint satisfaction mentioned on Slide 10?

Q4. The guidance function is constructed for three objectives, but I’m particularly concerned about the computational complexity of the reachability guidance. It seems to require numerous calculations and constraint functions. With such a large combination of functions, I wonder if the model can effectively learn and if the training time would be excessively long. Did the paper propose any specific training strategies or complementary techniques to address this? While the Author Insight section hints at this, honestly, it makes me wonder how it was trained at all!

정승재_teclados
Автор

Your presentation was very impressive, and it was interesting that it could create indoor scenes with physically plausible interactions.

I have two questions:
Q1. You said that the 3D-FRONT dataset is used for training and the GAPartNet dataset is used for inference. How is the difference between objects in the two datasets compensated for? Even if a similar object is selected, some position differences are expected to exist.

Q2. What exactly is the difference between Object Collision Rate and Scene Collision Ratio metrics?

Thank you!

tjswodud-cc
Автор

Thank you! It was a very insightful seminar 😊

I have two questions.

Q1. How did you address the mismatch issues when combining the static dataset (3D-FRONT) with the interactive dataset (PartNet)?

Q2. For articulated objects, does the model consider the full range of motion when generating the layout?

misong-kim