Why I Want a Gradient Camera

Think of image intensity as a height field and the analog-to-digital converter as a ruler. Its range depends on the number of quantization levels and their spacing. An 8-bit converter has 256 levels. For the six-log-unit dynamic range used in the example, a linear intensity measurement needs steps of 10⁶/256, losing fine detail.

A logarithmic measurement spans the same range with steps of 6/256 ≈ 0.0234 log units. The original example illustrates how quantization can still erase fine structure.

The gradient camera measures differences of log intensities. The converter’s reference follows a neighboring pixel rather than being tied to zero. Its range can therefore be devoted to local contrast.

The original discussion uses a 2% contrast-change example: log₁₀(1.02) ≈ 0.0086. Across 256 steps, this represents about 2.2 log units, or an intensity ratio of 1.02²⁵⁶ ≈ 159.09. This is a range for neighboring-pixel differences, not for absolute image intensity.

Natural-image gradients tend to cluster near zero. Summing local differences reconstructs a one-dimensional signal; in two dimensions, a Poisson solver integrates the measured gradient field.

The key relation is log(KI(x + 1)) − log(KI(x)) = log(I(x + 1)) − log(I(x)). A shared local gain K cancels from the measurement and does not have to be stored.

This differs from designs that subtract an adaptation value from the intensity signal and must retain that value to recover low-frequency structure. Gradient measurements can preserve both low- and high-frequency image information through integration.

Very large gradients can still saturate. However, the original analysis identifies fewer saturated differences than saturated absolute intensities, and many can be recovered using integrability: the integral around a closed loop must be zero. These constraints allow local recovery of missing gradient measurements.

We propose a camera that measures static gradients instead of static intensities. Quantizing sensed intensity differences between adjacent pixel values permits an ordinary A/D converter to measure detailed high contrast (HDR) scenes. We measure alternating ‘cliques’ of sensors (small groups) that locally determine their own best exposure, and reconstruct the image using a Poisson solver. This intrinsically differential design suppresses common-mode noise, hides and smoothes quantization, and can correct for its own saturated sensors. Simulations demonstrate these capabilities in side-by-side comparisons.