• **DMD** — micromirror array, each mirror tilts ± → **binary amplitude**, very fast (kHz); DLP projectors, single-pixel / compressive imaging, coded exposure & illumination

• **LCD / LCoS** — liquid-crystal pixels rotate polarization → with a polarizer give **greyscale amplitude**; in **phase-only** mode the controllable retardance imposes an arbitrary **phase** profile (programmable lens / hologram / aberration corrector)

• a programmable SLM is the **active** cousin of the fixed coded masks above (and of metalenses) — same idea (shape the wavefront), but reconfigurable on the fly

• **camera arrays** (Wilburn et al. 2005): a **grid of synchronized cameras** captures many viewpoints at once. Fuse them for **synthetic aperture** (below), **high-speed** video (stagger the triggers → interleave frames), **high dynamic range** (vary exposure per camera), or **super-resolution** (sub-pixel-shifted full-res views). The Stanford **multi-camera array** is the canonical research instrument.

• **synthetic aperture**: align all the small-aperture array views to a chosen **focal plane** and **average** them. The result behaves like a **single lens with an aperture as wide as the whole array** — an extremely shallow depth of field that **blurs foreground occluders into oblivion**, letting you **see through** foliage/crowds to a subject behind them; refocus by choosing a different alignment plane. The array's signature trick.

• **multi-camera phones**: modern phones carry a **cluster of cameras of different focal lengths** — **ultrawide · wide · tele** (sometimes a periscope tele) — and **fuse** them: switch/blend between modules across the zoom range, borrow detail from the sharper module, and use the **disparity between cameras** for **depth / portrait-mode** background blur. Note the caveat: this is **discrete optical steps + computational interpolation**, *not* a continuous optical zoom — and parallax between modules must be reconciled in software.

• **light.co** — the (now-defunct) **Light L16**: sixteen small camera modules of mixed focal lengths in a phone-sized body, fused into one high-resolution image with computed depth — a commercial **multi-aperture array** product; instructive as both the promise and the difficulty of array fusion.

• **bullet time**: a **ring/arc of synchronized cameras** fired at once (or in rapid sequence) → a **frozen instant viewed from a sweeping viewpoint** (the *Matrix* effect). A camera array used for its **angular** dimension rather than its aperture — the most recognizable array demo.

• **plenoptic vs array — the tradeoff to land**: both sample the light field, but a **plenoptic** camera buys **fine angular** sampling at the cost of **spatial** resolution (angular samples come *out of* the sensor's pixels), while an **array** keeps each view at **full resolution** but samples angle **coarsely** over a **wide baseline** (→ stronger parallax, true synthetic aperture). Form factor (one body vs a rig) and baseline (mm vs cm–m) are the practical axes.

• plus non uniformity of parameterization, paraxial

• **integral photography (Gabriel Lippmann, 1908)** — the origin of light-field capture: photograph the scene through an **array of tiny lenses** (a "fly's-eye" / lenslet sheet) onto a single plate, so each lenslet records a slightly **different view** (an **elemental image**) → the plate stores the scene's **directional** information, i.e. (a slice of) the light field. Viewed back through the *same* lens array it replays a **3-D, parallax-correct (autostereoscopic)** image. The direct ancestor of every plenoptic camera and of lenticular / **integral-imaging 3-D displays**.

• **the plenoptic / light-field camera** lineage: **Adelson & Wang 1992** (the plenoptic camera) → **Ng 2006 / Lytro** — a **microlens array** in front of the sensor samples the rays *inside* the main lens (the same integral-photography trick, now digital), trading **spatial resolution for angular samples** → **refocus, viewpoint shift, and depth after capture** (the "capture the full set, decide later" lesson, L14).

• **Lytro** (and **Raytrix**) as products; cross-ref the Intro history (Lippmann; Lytro 2012).

• **shift-and-add refocusing** — once the light field is captured, choosing a focus plane is just **summing the sub-aperture views, each shifted in proportion to its position in the aperture** (a shear of the light field, then integrate over angle). Different shifts pick different virtual focus planes, so a *single* capture yields a whole **focal stack**, refocusable **after** the shot. Ng's **Fourier-slice photography** (2005) is the frequency-domain fast version (refocusing = taking a sheared 2-D slice of the 4-D light field's spectrum).

• this is the **dual of focus stacking** ([[Multiple exposure imaging]] — *Focal stacks and depth of field extension*): there you capture **frames** at stepped focus and merge; here you capture **rays** and *synthesise* the frames, then the **same sharpness-weighted merge** (with the high exponent) collapses them to all-in-focus. The per-pixel sharpest plane is again a coarse **depth map** (shape-from-focus). Both are **L14** — capture the full set, decide later — on the focus axis.

see my blog https://www.thecomputationalphotographer.net/2020/02/i-want-a-light-field-camera/

• tracking

• and traps

• mount a light on a **quadrotor** and let it *fly itself* into position: a feedback loop measures the subject's **rim light** (the bright contour that separates subject from background) and commands the drone to hold a target rim width/placement as the subject or photographer moves — the light source becomes an **autonomous, computationally-controlled actuator** rather than a fixed stand. Computational illumination where the optimized degree of freedom is the **light's position in space**; ties the lighting sections above to the robotics/active-capture theme (cross-ref FUNDAMENTALS — *Cameras beyond photography*).

• sidebar: Yang privacy

• time lapse tricks

• not perfect place but OK

• a taxonomy of ways to recover per-pixel **range $Z$** — the dimension a single camera throws away in the divide-by-$Z$ projection (FUNDAMENTALS):

• **stereo** (passive): two cameras a baseline $B$ apart; triangulate depth from **disparity**, $Z = fB/d$ (cross-ref *Multiple view geometry*). Cheap and passive, but fails on **textureless / repetitive** surfaces — the correspondence problem.

• **stereo with structured light** (active): *project* a known pattern — stripes, a pseudo-random **speckle / dot** pattern, or phase-shifted sinusoids — so correspondence is solved even on a blank wall; one camera + one projector triangulate. The original **Kinect v1** (PrimeSense speckle); industrial **phase-shift profilometry** for 3-D scanning. Robust indoors; struggles in bright sun and on specular / very dark surfaces.

• **LiDAR** (light detection and ranging): emit a laser pulse and time its return (**direct time-of-flight**) for each scanned or flashed direction → a depth **point cloud**. Scanning (spinning or MEMS-mirror) vs **flash** LiDAR (a SPAD array — cross-ref single-photon sensors below); long range, works in the dark — the basis of self-driving range sensing and recent phone/tablet LiDAR.

• **time-of-flight** (the other active route): continuous-wave (**phase**) ToF — e.g. **Kinect v2** — vs direct/pulsed ToF (LiDAR); see the ToF section below.

• others, in passing: depth from **focus / defocus** and **dual-pixel** (Optics — Focus/AF), **monocular learned depth** (COMPUTATIONAL TOOLS — Deep learning), and shape-from-X.

• per-pixel **velocity** from a frequency / phase shift — optical Doppler, Doppler **lidar / radar**, laser **vibrometry**; the imaging cousin of medical **Doppler ultrasound** and Doppler weather radar

• measures motion **directly** (radial velocity) rather than inferring it from frame-to-frame optical flow → fast, robust in low light; pairs with ToF (range + range-rate)