Quantization

• Lossless: If all numbers in a vector are below 255: use bytes to encode the vector.
• Lossy: Reducing 64-bit floats to 32-bit floats.

Quadtree

• Splitting the height map evenly in to four smaller planes.
• Evaluate which quantization method would give the smallest size without exceeding error tolerance
• Include the cost of overhead for each of the four splits--sometimes its better to just not bother splitting.

Delta encoding

• Encode pixels as relative changes from other nearby pixels.
• Changes between neighbors can often be small enough that the data can now be quantized effectively.

Regression

• Compute a line/curve for a strip of pixels and represent the data with it
• Store residuals from the curve (which should now be much smaller and easier to quantize.)