Implicit Neural Representations: From Objects to 3D Scenes

Explore a keynote presentation on implicit neural representations for 3D scene reconstruction and understanding. Delve into advanced techniques for overcoming limitations of fully-connected network architectures in implicit approaches. Learn about a hybrid model combining neural implicit shape representation with 2D/3D convolutions for detailed object and large-scale scene reconstruction. Discover methods for capturing and manipulating visual appearance through surface light field representations. Gain insights into recent efforts in collecting real-world material information for training these models. Examine the KITTI-360 dataset, featuring 360-degree sensor data and semantic annotations for outdoor environments. Cover topics including convolutional occupancy networks, object-level and scene-level reconstruction, appearance prediction, generative models, and joint estimation of pose, geometry, and SVBRDF.

Intro
Collaborators
3D Representations
Limitations
Convolutional Occupancy Networks
Comparison
Object-Level Reconstruction
Training Speed
Scene-Level Reconstruction
Large-Scale Reconstruction
Key Insights
Problem Definition
Existing Representation
Overfitting to Single Objects
Single Object Experiments
Single Image Appearance Prediction
Single View Appearance Prediction
Generative Model
Materials
Joint Estimation of Pose, Geometry and SVBRDF
Qualitative Results
3D Annotations