Depth Amplification: Mechanisms in Imaging

• Depth Amplification Effect is a collection of mechanisms that enhance the observable impact of depth variations in imaging systems through scaling depth-related blur, geometry, and focal properties.
• Key techniques include render-time amplification in AR-GANs, parallax-driven synthesis, and focus sweeps which improve depth inference and synthetic defocus control.
• The concept also covers axial-range extension and digital compensation methods that mitigate depth-dependent degradation and refine imaging accuracy.

Depth Amplification Effect is not a standardized term in the literature represented here. The closest recurring idea is that a system makes axial or geometric differences more consequential in observation, rendering, or inference: by scaling disparity before aperture integration, turning handheld parallax into synthetic defocus, refining relative depth so that relit normals become more plausible, sweeping focus so that each depth layer traces a characteristic blur trajectory, digitally compensating depth-dependent signal loss, or, conversely, allowing depth-dependent blur to amplify measurement uncertainty. This suggests that the term is best treated as an umbrella label for several non-equivalent mechanisms rather than a single formal effect (Kaneko, 2021, Wu et al., 2022, Bhattarai et al., 19 Dec 2025, Kim et al., 2016, Boroomand et al., 2015, Rajendran et al., 2019).

1. Terminological status and conceptual scope

Several relevant papers explicitly do not use the exact phrase “Depth Amplification Effect.” In "Unsupervised Learning of Depth and Depth-of-Field Effect from Natural Images with Aperture Rendering Generative Adversarial Networks" (Kaneko, 2021), the closest mechanisms are a depth scaling factor $s$, a center-focus depth gain $g$, and instance-dependent scaling by $10 \times \mathrm{MLP}(z)$. In "Re-Depth Anything: Test-Time Depth Refinement via Self-Supervised Re-lighting" (Bhattarai et al., 19 Dec 2025), the closest phenomenon is refinement that makes depth “more geometrically expressive” rather than a scalar gain. In "Video Depth-From-Defocus" (Kim et al., 2016), the relevant effect is that controlled focus variation makes depth-dependent blur more observable over time. In "Diffractive flat lens enables Extreme Depth-of-focus Imaging" (Banerji et al., 2019), the paper is explicit that the device amplifies usable axial imaging range, not scene-depth remapping. In "Depth Compensated Spectral Domain Optical Coherence Tomography via Digital Compensation" (Boroomand et al., 2015), the nearest term is “depth compensation.”

This suggests four distinct interpretations. First, render-time amplification increases the effective mapping from estimated depth or disparity to visible blur. Second, observability amplification makes depth easier to estimate by causing depth layers to evolve differently across viewpoints, focus settings, or illumination conditions. Third, axial-range amplification enlarges the interval over which an imaging system remains useful or restores information that would otherwise degrade with depth. Fourth, uncertainty amplification has the opposite sign: depth-related optics increase estimation error rather than depth salience. Treating these as interchangeable is a common source of confusion.

2. Aperture rendering and depth-to-blur gain

AR-GANs provide the clearest rendering-side analogue of a depth amplification mechanism. The generator is decomposed as

$I_d^g = G_I(z), \qquad D^g = G_D(z),$

with $G_I$ and $G_D$ sharing all weights except the last layer. The generated scalar field is explicitly interpreted as an aperture-warping disparity-like quantity because the paper states that “depth and disparity [are used] interchangeably to indicate disparity across a camera aperture.” A physically motivated aperture renderer then maps $(I_d^g,D^g)$ to a shallow-DoF image through

$M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$

$$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$

$$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$

with the focal plane at $g$0 and out-of-focus regions at $g$1. In practical terms, larger offsets $g$2 yield stronger defocus, so any mechanism that enlarges $g$3 or $g$4 strengthens blur (Kaneko, 2021).

The main control variable is DoF mixture learning: $g$5 Here $g$6. When $g$7, the renderer outputs a deep-DoF image close to $g$8; when $g$9, it outputs the shallow-DoF image $10 \times \mathrm{MLP}(z)$0. The appendix also studies continuous sampling with $10 \times \mathrm{MLP}(z)$1, where $10 \times \mathrm{MLP}(z)$2 and $10 \times \mathrm{MLP}(z)$3 become the deepest- and most shallow-DoF endpoints. The implementation further rescales depth in the generator head as

$10 \times \mathrm{MLP}(z)$4

A plausible implication is that AR-GAN already contains an implicit depth-gain stack: latent-dependent magnitude scaling, linear scaling before rendering, and the aperture integral itself jointly determine how strongly predicted depth appears as blur.

The center-focus prior adds an explicit depth gain,

$10 \times \mathrm{MLP}(z)$5

The paper sets $10 \times \mathrm{MLP}(z)$6, $10 \times \mathrm{MLP}(z)$7 for $10 \times \mathrm{MLP}(z)$8 training, and $10 \times \mathrm{MLP}(z)$9 for $I_d^g = G_I(z), \qquad D^g = G_D(z),$0. This prior is used only at the beginning of training because keeping it until the end greatly reduces DSD. Experimentally, the appendix uses DSD as a proxy for meaningful depth spread; on Oxford Flowers, DSD rises from $I_d^g = G_I(z), \qquad D^g = G_D(z),$1 for “$I_d^g = G_I(z), \qquad D^g = G_D(z),$2 only” and $I_d^g = G_I(z), \qquad D^g = G_D(z),$3 for “$I_d^g = G_I(z), \qquad D^g = G_D(z),$4 only” to $I_d^g = G_I(z), \qquad D^g = G_D(z),$5 for mixture training with $I_d^g = G_I(z), \qquad D^g = G_D(z),$6. The paper does not study $I_d^g = G_I(z), \qquad D^g = G_D(z),$7, but it explicitly notes that blur strength is controlled through the scaling of $I_d^g = G_I(z), \qquad D^g = G_D(z),$8, making extrapolative amplification an evident but untested extension (Kaneko, 2021).

3. Parallax-driven and relighting-driven amplification of geometry

"Direct Handheld Burst Imaging to Simulated Defocus" (Wu et al., 2022) replaces explicit depth-to-blur rendering with direct exploitation of burst parallax. The central statement is that “the simulated aperture size equals the user’s lateral hand translation during burst acquisition” and “the simulated defocus blur for each region is as strong as its disparity.” After homography-based orientation correction and image-space refocusing, the residual inter-frame motion is depth-dependent parallax. The supervision equation is the standard synthetic-aperture integral

$I_d^g = G_I(z), \qquad D^g = G_D(z),$9

with a target formed from 49 viewpoints inside a circular aperture from a $G_I$0 light field, while the input is a sparse 9-viewpoint handheld trajectory. The architecture uses a Blur Prediction Network and a Multi-scale Merging Network; large disparities are handled at coarser scales because they shrink in pixel units. Quantitatively, on the Stanford Lytro Light Field dataset, the paper reports SSIM/LPIPS of $G_I$1 for a 4-frame burst and $G_I$2 for a 9-frame burst, compared with $G_I$3 for the center view only and $G_I$4 for the comparison method using 4 corner viewpoints. In this formulation, the effective “amplification” is not a depth-map gain but the conversion of small handheld translations into a larger synthetic aperture whose blur strength grows with residual disparity (Wu et al., 2022).

"Re-Depth Anything: Test-Time Depth Refinement via Self-Supervised Re-lighting" (Bhattarai et al., 19 Dec 2025) operates differently. It starts from DA-V2 disparity, converts it to depth, derives normals by unprojecting into 3D and computing

$G_I$5

then re-lights the input with a Blinn-Phong model,

$G_I$6

and optimizes only intermediate embeddings and the decoder under SDS and a disparity regularizer. The paper is explicit that shading is largely insensitive to global depth scale; what changes is local and meso-scale structure, surface orientation coherence, and relative depth ordering. On KITTI, $G_I$7 improves from $G_I$8 to $G_I$9, AbsRel from $G_D$0 to $G_D$1, and SI log from $G_D$2 to $G_D$3. On ETH3D, $G_D$4 improves from $G_D$5 to $G_D$6, AbsRel from $G_D$7 to $G_D$8, and RMSE from $G_D$9 to $(I_d^g,D^g)$0. The qualitative claim is not scalar depth magnification but stronger local relief, corrected semantic shape bias, sharper boundaries, and removal of implausible detail. A plausible reading is that the method amplifies the structural expressiveness of depth rather than its numeric range (Bhattarai et al., 19 Dec 2025).

4. Focus sweeps, focal stacks, and the amplification of depth observability

"Video Depth-From-Defocus" (Kim et al., 2016) is a direct formulation of measurement-side depth amplification. The input is a video in which the focus plane is continuously moved back and forth, producing focus ramps. Under the thin-lens model, the circle of confusion is

$(I_d^g,D^g)$1

so, with fixed $(I_d^g,D^g)$2 and $(I_d^g,D^g)$3, blur depends on scene depth $(I_d^g,D^g)$4 and framewise focus distance $(I_d^g,D^g)$5. As $(I_d^g,D^g)$6 sweeps, each depth layer generates a characteristic temporal blur trajectory. The method jointly estimates per-frame depth maps $(I_d^g,D^g)$7, all-in-focus frames $(I_d^g,D^g)$8, and focus distances $(I_d^g,D^g)$9, using a data term, spatial and temporal smoothness, and numerical focus refinement with $M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$0. The optimization is embedded in a 3-level pyramid with 3 iterations per level. This is the clearest case in which “amplification” refers to making depth more observable by inducing controlled time-varying defocus; the same recovered RGB-D video is then used for synthetic refocusing, increased aperture blur, and tilt-shift effects (Kim et al., 2016).

"The Application of Preconditioned Alternating Direction Method of Multipliers in Depth from Focal Stack" (Javidnia et al., 2017) addresses a related but narrower problem: recovering a better depth map from a focal stack so that synthetic defocus is cleaner. The focus measure is the Modified Laplacian,

$M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$1

and the per-pixel focus function is modeled by a 3-point Gaussian,

$M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$2

The paper then refines depth with a PADMM-based variational solver. On 21 focal-stack datasets, the reported PADMM setting uses a maximum of 300 iterations and a regularization parameter of 0.7; convergence is reported around iteration 226, where the decay-of-energy level is about 0.01, while the compared method is at about 3.6. The total pipeline time on $M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$3 images is about 53 s, with about 1.5 s for the modified PADMM optimization. This is not an explicit amplification method: it improves structural accuracy, corners, edges, and occlusion handling so that later depth-aware defocus can be stronger without boundary leakage or mixed layers (Javidnia et al., 2017).

5. Axial-range amplification and depth compensation in physical optics

In flat optics, "Diffractive flat lens enables Extreme Depth-of-focus Imaging" (Banerji et al., 2019) uses the term in an axial-range sense rather than a scene-geometry sense. The conventional scaling is

$M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$4

The paper designs a multi-level diffractive lens by gradient-descent-assisted binary search over an axial interval with $M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$5, $M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$6, $M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$7, aperture $M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$8, minimum ring width $M(\mathbf{x}, \mathbf{u}) = T(D(\mathbf{x})),$9, at most 100 height levels, and maximum height $$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$0. The reported conventional DOF at the highest NA is $$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$1, while the achieved axial span yields

$$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$2

described as “over 4 orders of magnitude” enhancement. Experiments report focus from $$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$3 to about $$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$4, measured FWHM from $$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$5 at $$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$6 to $$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$7 at $$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$8, and a $$L(\mathbf{x}, \mathbf{u}) = I_d(\mathbf{x} + \mathbf{u} M(\mathbf{x}, \mathbf{u})),$$9 field of view. The paper is explicit that this is not depth remapping, stereoscopic disparity enhancement, or nonlinear expansion of scene depth; it is an extended, nearly invariant focal response (Banerji et al., 2019).

In OCT, "Depth Compensated Spectral Domain Optical Coherence Tomography via Digital Compensation" (Boroomand et al., 2015) treats the relevant effect as restoration of deep signal quality. The observation model is

$$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$0

with depth-dependent axial PSF $$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$1, lateral PSF $$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$2, sidelobe term $$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$3, and multiplicative speckle $$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$4. The compensated tomogram is estimated by MAP inference,

$$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$5

with a stochastically fully-connected CRF prior. The reported outcome is an average axial resolution improvement of about $$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$6, an average lateral resolution improvement of about $$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$7, and an average SNR improvement of about $$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$8, with the compensated SNR curve becoming almost uniform with depth. This is best understood as digitally compensating depth-dependent degradation rather than magnifying metric depth itself (Boroomand et al., 2015).

6. Uncertainty amplification and the boundaries of the term

"Uncertainty amplification due to density/refractive-index gradients in volumetric PTV and BOS experiments" (Rajendran et al., 2019) gives the most explicit negative formulation of a depth amplification effect. The image of a particle or dot is modeled as a sum of diffraction-limited ray images, and the apparent displacement of ray $$I_s(\mathbf{x}) = \sum_{\mathbf{u}} A(\mathbf{u}) L(\mathbf{x}, \mathbf{u}),$$9 scales as

$g$00

After moment matching, the effective image widths become

$g$01

$g$02

Thus deeper particles, larger target-to-flow distances, stronger nonlinear density gradients, and wider ray cones increase blur, which in turn raises the CRLB for centroid estimation. The paper packages this effect into an Amplification Ratio for BOS, defined from the ratio of centroid uncertainties between gradient and reference images. Here “amplification” means amplification of error, not enhancement of geometric representation (Rajendran et al., 2019).

Outside imaging, related terminology diverges further. "Topological directed amplification" (Midya, 2022) studies transient norm growth in asymptotically stable nonnormal skin-effect lattices, with onset governed by $g$03; the relevant “depth” is propagation time or distance in a lattice, not scene depth. "Beyond Worst-Case Branching: Quantum Tree Search via Amplitude Amplification" (Wichert, 26 Jun 2026) uses amplitude amplification in the Grover sense, with search cost $g$04 or $g$05; the depth variable $g$06 is tree depth. "Hardness Amplification and the Approximate Degree of Constant-Depth Circuits" (Bun et al., 2013) proves hardness amplification under $g$07-composition, where increasing circuit depth by one layer strengthens approximation hardness. These are legitimate amplification phenomena, but they are not the same object as depth-dependent blur, focus, or axial imaging effects. The phrase therefore has no single field-independent meaning.

A consistent synthesis is possible only at a high level. In imaging, the most coherent use of “Depth Amplification Effect” refers to any mechanism that increases the operational consequences of depth or axial variation—by making blur stronger, geometry more inferable, axial tolerance wider, or, in adverse cases, measurement uncertainty larger. The literature does not yet define a single canonical effect by that name; it instead provides a family of mechanisms with different signs, different observables, and different physical meanings.