Instant-NGP Hash Encoding Quiz

Answers and explanations appear as you take the quiz.

1. What is the main bottleneck in classic NeRFs that Instant-NGP targets?
   • The large MLP queried millions of times per scene
   • Network bandwidth for streaming
   • The display refresh rate
   • Audio synchronization
2. Where does Instant-NGP store most of its learnable parameters?
   • In a single dense voxel grid
   • In a multiresolution hash table of feature vectors
   • In the camera intrinsics
   • Only inside a very deep MLP
3. How are features combined across the multiple resolution levels?
   • They are averaged into one scalar
   • They are sorted alphabetically
   • They are looked up, trilinearly interpolated, then concatenated
   • They are discarded except the finest level
4. How does Instant-NGP handle hash collisions in its fine-resolution tables?
   • It forbids collisions entirely with a huge table
   • It crashes when collisions occur
   • It hashes only colors, never positions
   • It lets collisions happen and relies on training gradients to disambiguate
5. Roughly how much does Instant-NGP speed up NeRF training?
   • From hours down to seconds (orders of magnitude)
   • It is actually slower
   • About 5 percent faster
   • No change, only better quality
6. Beyond NeRFs, what else can hash-grid encoding represent?
   • Only printed text
   • Signed distance functions and gigapixel images
   • Only stock charts
   • Only audio waveforms
7. Why does Instant-NGP allow different fine-grid cells to share a hash-table entry?
   • To make every spatial coordinate map to a unique table entry
   • To remove the need for trainable feature vectors
   • To keep the feature tables bounded in memory while representing higher-resolution spatial detail
   • To replace all grid levels with one full-resolution dense array
8. How do Instant-NGP's resolution levels balance collisions and spatial detail?
   • Every level uses one dense grid with no feature table
   • Coarse levels share all entries while fine levels are guaranteed collision-free
   • Only the highest-resolution level stores learnable features
   • Coarse levels can be collision-free while finer levels share entries to represent more detail within bounded memory