Instant NeRF

• Instant Neural Graphics Primitives with a Multiresolution Hash Encoding
• Neural Radiance Field
• Neural graphics primitives, parameterized by fully connected neural networks, can be costly to train and evaluate
• rely on task specific Data Structures
• new input encoding that permits the use of a smaller network without sacrificing quality
• educing the number of floating point and memory access operations
• near-instant training of neural graphics primitives on a single GPU for multiple tasks
• small neural network is augmented by a multiresolution hash table of trainable feature vectors whose values are optimized through Gradient Descent
• automatically focuses on relevant detail, independent of task at hand
• low overhead
• In a gigapixel image, they represent an image by a neural network. SDF learns a signed distance function in 3D space whose zero level-set represents a 2D surface
• 2D images and their camera poses to reconstruct a volumetric radiance-and-Density field that is visualized using ray marching.
• neural volume learns a denoised radiance and Density field directly from a volumetric path tracer.
• only vary the hash table size which trades off quality and performance
• disambiguate hash collisions, making for a simple architecture that is trivial to parallelize on modern GPUs
• parallelism
• fully-fused Operator Fusion CUDA kernels with a focus on minimizing wasted bandwidth and compute operations