Computational photography

The workbench groups the experiments by the operation that unlocks them: lift image values into a higher-dimensional grid, vote in parameter space, align observations despite bad matches, or integrate measurements across exposure and viewpoint. Each group makes an intermediate representation inspectable.

This collection places an algorithm's intermediate representation beside its input and output. A bilateral-grid view shows image samples arranged by position and intensity. Hough views show how evidence in image space becomes votes for a geometric explanation. Controls connect the representation's resolution and thresholds to the visible result.

The grid implementation stores RGB values and occupancy weights in a Float32Array. Pixels are assigned to spatial and intensity bins, and output values are sampled through trilinear interpolation. The accompanying three-dimensional view uses cell color and weight to expose how samples populate the grid. Its value is the ability to inspect the discretization and the result together.

For lines, each selected image pixel votes in a discretized rho–theta accumulator. Sine and cosine values are precomputed for the angular bins, keeping repeated trigonometric evaluation outside the pixel loop. A neighborhood maximum check identifies candidate peaks, which are ranked by vote count and drawn back into the original image.

The generalized Hough study builds an R-table from gradient-direction bins and vectors relative to a drawn template's centroid. It explores candidate positions over sampled scales and rotations. Template and target sampling reduce the amount of voting, making the cost of adding transformation dimensions explicit. Circle views similarly expose radius-dependent accumulators rather than presenting only a final detection overlay.

Begin with a few drawn lines and compare the image, accumulator, and detected geometry. Raise the vote threshold to see which structures survive, then use the line-field display to examine how parameter-space evidence maps back into image space. In the grid view, vary spatial scale and intensity spacing to inspect changes in the occupied cells.

Each study links its numerical data, candidate selection and visual display, so an intermediate result can help explain the output.

A scene rendered by Aperture supplies linear radiance before display mapping. Simulated camera exposures turn that HDR source into clipped, noisy brackets; inspect their contributions, recover radiance, then choose how to display it. Exposure spacing changes which highlights and shadows are measurable.

The input is an actual path-traced render with retained HDR values, and the exposure brackets use a known camera response. This isolates calibrated radiance merging; unknown-response calibration, motion alignment and deghosting are separate problems.

A simulated 10 × 10 camera array renders a three-dimensional scene from 100 positions. The cameras share orientation and an off-axis reference plane through the common target, so their recorded rays have a consistent calibration. Each view includes its own occlusions, reflections and refractions.

Choose a new focus plane and shear the images in both camera-array dimensions, then average the selected aperture in linear light. Change aperture size to trade depth of field against angular support. A row of cameras supplies the epipolar view; the full two-dimensional array produces the refocused image.

The camera-array images are rendered ahead of time with Aperture; browser refocusing uses the captured images alone. LightField / Neural extends this pipeline with traced scene capture, broader ray-space inspection and neural-representation comparisons.

Two overlapping crops of the same rendered scene form a small stitching experiment. Supplied correspondences include controlled noise and false matches. Least squares fits every match; RANSAC selects a consensus, then refits. Inspect the match residuals and compare the resulting stitched panoramas, seam and overlap differences.

The experiment uses a scale/rotation/translation model with known source correspondences. It makes robust fitting and image warping visible without adding feature detection to the same lesson. Arbitrary camera perspective needs a richer geometric model, and parallax may need more than one warp.

In gradient space, a pasted patch becomes a boundary-constrained reconstruction. Compare a hard paste against source-gradient and mixed-gradient Poisson solutions; inspect the derivative field and solver residual while moving the patch.

In Fourier space, inspect the centered spectrum and edit a low-pass, high-pass, band or conjugate notch mask. The inverse transform and signed removed component show exactly what the mask kept or discarded. Soft boundaries trade sharp selection for less ringing.

The frequency-editing sample is a viewport capture from RadianceLab, showing glass and a caustic. Its displayed RGB values are appropriate for these image operations. Aperture supplies the separate linear HDR fixture and rendered light-field array; the two sources are credited where they appear.