Towaki Takikawa

A foundational paper on scaling up 3D generation with smarter data structures for high-resolution text-to-3D content creation.

Developed a framework for amortized text-to-3D object synthesis, allowing knowledge sharing between prompts for faster convergence and smoother interpolations in 3D asset generation.

Introduced a method combining learned hash probing with hash tables for more compression and faster processing of neural graphics primitives.

Explored learned indexing for more compressive neural field representations than hash tables and grids, enabling significant memory reduction and multiresolution representation.

Co-authored a paper on RTMV, a ray-traced multi-view synthetic dataset for novel view synthesis.

Proposed differentiable data structures, neural fields, and GPU programming for fast, real-time rendering of implicit 3D shapes, achieving high fidelity and efficiency.

Introduced a two-stream CNN architecture for semantic segmentation, explicitly wiring shape information for better computer vision by paying attention to object boundaries.