DepthDark: Techniques for Low-Light Depth Recovery

Updated 7 July 2026

DepthDark is a research motif encompassing various methods to estimate depth in settings where visible-light cues fail, including low-light, underwater, and defocus scenarios.
These approaches employ techniques such as robust median filtering, adversarial regularization, synthetic flare and noise simulation, and uncertainty-driven fusion.
The methods enhance applications like visual mapping, 3D reconstruction, and depth completion by integrating diverse modalities, including LiDAR and thermal imaging.

to=arxiv_search.search 񹚊ppjson
{"query":"DepthDark", "max_results": 10}analysis to=arxiv_search.search code 大发极速json
{"query":"\"Robustly Removing Deep Sea Lighting Effects for Visual Mapping of Abyssal Plains\"", "max_results": 5}analysis to=arxiv_search.search code 一级a做爰片json
{"query":"\"Regularizing Nighttime Weirdness: Efficient Self-supervised Monocular Depth Estimation in the Dark\"", "max_results": 5}analysis to=arxiv_search.search code 皇轩json
{"query":"\"DepthDark: Robust Monocular Depth Estimation for Low-Light Environments\"", "max_results": 5}analysis to=arxiv_search.search code 天天中彩票足球 to=arxiv_search.search analysis code 天天彩票软件json
{"query":"\"Dark3R: Learning Structure from Motion in the Dark\"", "max_results": 5}analysis to=arxiv_search.search code 北京赛车有json
{"query":"\"All-day Depth Completion\"", "max_results": 5}analysis to=arxiv_search.search code 天天爱彩票 to=arxiv_search.search analysis code уйғурларjson
{"query":"\"DarkVGGT: Seeing Through Darkness Using Thermal Geometry without Daylight Tax\"", "max_results": 5}analysis to=arxiv_search.search code 天天中彩票在json
{"query":"\"Dark Channel-Assisted Depth-from-Defocus from a Single Image\"", "max_results": 5}analysis to=arxiv_search.search code
DepthDark is a non-canonical designation that has been applied to several technically distinct methods concerned with perception under darkness, low illumination, or illumination corruption. In the cited literature, the name appears in at least five computer-vision settings: parameter-free removal of co-moving lighting artefacts in deep-sea imagery; self-supervised nighttime monocular depth estimation; parameter-efficient adaptation of a foundation model for low-light monocular depth estimation; dark-channel-assisted single-image depth-from-defocus; and an all-day depth-completion pipeline that combines sparse depth with imagery. Closely related work also includes darkness-robust structure from motion and RGB-T feed-forward geometry, which address the same failure modes of RGB-only geometry in low-SNR or low-visibility regimes [2110.00480], [2108.03830], [2507.18243], [2506.06643], [2405.17315], [2603.05330], [2606.11326].

1. Terminological scope and research context

The term “DepthDark” does not denote a single standardized architecture. Instead, it labels a family of approaches whose common concern is geometry or image formation under conditions where visible-light assumptions fail. In the deep ocean, the dominant issue is not ambient darkness alone but robot-borne illumination, backscatter, attenuation, and moving light cones, which corrupt visual mapping and surface-albedo estimation [2110.00480]. In nighttime driving and general low-light scenes, the dominant issues are low visibility, weak texture, weak or absent edges, flickering or spatially varying illumination, flare, and sensor noise, all of which destabilize monocular depth learning or inference [2108.03830], [2507.18243].

A further branch uses the label for single-image depth-from-defocus. There, darkness enters through the dark channel prior rather than by explicit nighttime capture: local dark-channel statistics are used as a complementary cue to infer depth from a single defocused RGB image [2506.06643]. Another branch addresses depth completion under day and night by combining an image with synchronized sparse depth from LiDAR, using uncertainty to determine when the sparse modality should dominate in poorly illuminated regions [2405.17315].

This suggests that “DepthDark” is best understood as a research motif rather than a unique method name: a recurring attempt to recover geometry, albedo, or dense depth when standard RGB photometric cues become unreliable.

2. Deep-sea imaging formulation and the original parameter-free compensation paradigm

In “Robustly Removing Deep Sea Lighting Effects for Visual Mapping of Abyssal Plains” [2110.00480], the camera image at pixel $x$ and pose $P$ is modeled as the sum of a multiplicative direct-return term and an additive scatter term. The classical underwater model is written as
[
I_c(x,P)
\;=\;
A_s(x,P)\,\bigl[\,I_l(\theta,\phi)\,e^{{-\eta\,d_l(x,P)}\,e^{{-\eta\,d_c(x,P)}/d^2(x,P)\bigr]}}
\;+\;
I_s(x,P)\,.
]
By grouping geometry, lamp-pattern, and attenuation into a single factor image $F(x,P)$, and relabeling the scatter term as an additive image $B(x)$ under stationarity assumptions, the working model becomes
[
I_c(x,P)\;=\;A_s(x,P)\,F(x,P)\;+\;B(x)\,.
]

The method estimates both nuisance terms without explicit parameters for light, camera, water, or scene. The additive term is obtained from a brief high-altitude “water-only” sequence dominated by backscatter:
[
\widehat B(x)\;=\;\mathrm{median}\bigl{\,I_c(x,P_{\mathrm{high}^{{(i)}})\bigr}_{i=1\ldots} b}\,.
]
The multiplicative term is then obtained from a sliding temporal window of seafloor images. First an all-seafloor median is computed,
[
\widetilde I_{\mathrm{sf}}(x,P)\;=\;\mathrm{median}!\bigl{\,I_c(x,P^{{(j)})\bigr}_{j=-\lfloor} n/2\rfloor\ldots\lfloor n/2\rfloor},
]
and then converted to a factor image using a reference seafloor albedo $A_{\mathrm{ref}}$:
[
\widehat F(x,P)
\;=\;
\frac{\widetilde I_{\mathrm{sf}}(x,P)-\widehat B(x)}{A_{\mathrm{ref}}}\,.
]
The corrected image is
[
\widehat A_s(x,P)
\;=\;
\frac{\,I_c(x,P)-B(x)\,}{\,\widehat F(x,P)\,}\,.
]

The central statistical claim is robustness via pixel-wise medians. Because the median has a breakdown point of $50\%$, occasional bright or dark outliers do not spoil the backscatter estimate. For contamination fraction $c<0.5$, the failure probability for an $n$-sample median is
[
P_{\mathrm{fail}}
\;=\;
\sum_{u=\lceil n/2\rceil}ⁿ \binom{n}{u}\,(c)^{u(1-c)^{\,n-u}\,,}
]
which rapidly tends to zero as $n$ grows [2110.00480].

The full preprocessing pipeline consists of water-only capture, additive backscatter estimation, sliding-window all-seafloor median computation, factor-image estimation, pixel-wise compensation, and optional global white balance. The implementation requires no explicit modeling of lamp orientation or fall-off and bypasses distance estimation by collapsing unknown attenuation geometry into observed statistics. A CUDA prototype processes 12 MP images at approximately $2$ Hz on a single GPU. On real $4$ km-depth data, the method reportedly outperforms six state-of-the-art dehazing and enhancement methods by a no-ground-truth “color consistency error,” and its corrected images improve mosaics and 3D reconstruction quality [2110.00480].

3. Self-supervised nighttime monocular depth estimation

In “Regularizing Nighttime Weirdness: Efficient Self-supervised Monocular Depth Estimation in the Dark” [2108.03830], DepthDark is a self-supervised monocular depth framework with three named modules: Priors-Based Regularization (PBR), Mapping-Consistent Image Enhancement (MCIE), and Statistics-Based Masking (SBM). The formulation is driven by three nighttime failure modes: large textureless regions under low visibility, weak edges, and illumination variation that breaks brightness consistency.

PBR treats the depth network $\Phi_d$ as a generator in a PatchGAN-style adversarial game. Predicted and reference depths are normalized by
[
\mu(D)\; \triangleq\; D / \mathrm{mean}(D),
]
and the discriminator $\Phi_D$ receives the concatenation of a coordinate image $I_p$ and $\mu(D)$. The LSGAN losses are
[
L_D = \tfrac12 E_{D_r} [ (\Phi_D(\mathrm{cat}(I_p,\mu(D_r)))-1)² ]
+\tfrac12 E_{D_t} [ (\Phi_D(\mathrm{cat}(I_p,\mu(D_t))))² ],
]
[
L_G = \tfrac12 E_{D_t} [ (\Phi_D(\mathrm{cat}(I_p,\mu(D_t)))-1)² ].
]
This regularizer constrains local predicted-depth distributions to resemble unpaired reference depths [2108.03830].

MCIE uses a single shared histogram-equalization-style mapping $\gamma$ across target and source frames:
[
I't = \gamma(I_t), \qquad I'_s = \gamma(I_s).
]
The mapping is formed by clipping the brightness histogram, redistributing excess uniformly, computing the cumulative distribution function, and defining
[
\gamma(b) = \frac{\mathrm{cdf}(b) - \mathrm{cdf}{\min}}{\mathrm{cdf}{\max} - \mathrm{cdf}{\min}}\times(L-1),
]
with $L=256$. Because the mapping is global and shared, the enhanced photometric loss remains compatible with view synthesis:
[
L'_{pe}(I'_t,\hat I'_t) = (\alpha/2)(1-\mathrm{SSIM}(I'_t,\hat I'_t)) + (1-\alpha)|I'_t-\hat I'_t|_1,
]
with $\alpha=0.85$ [2108.03830].

SBM defines temporal intensity change
[
d_{ts} = |I_t-I_s|1,
]
tracks its running mean with EWMA,
[
\tilde d{ts}(i) = \beta \tilde d_{ts}(i-1) + (1-\beta)d_{ts}(i), \qquad \beta=0.98,
]
and removes unreliable pixels using a percentile threshold:
[
m_s(p) = [ d_{ts}(p) > p(\tilde d_{ts},\epsilon) ].
]
The final mask is
[
m = m_a \odot m_s.
]

The complete loss is
[
L_{\mathrm{total}} = \sum_{t,s} m \cdot L'_{pe} + \eta L_s + \xi L_G + \tau L_D,
]
with edge-aware smoothness
[
L_s = |\partial_x D_t| e^{{-|\partial_x} I_t|} + |\partial_y D_t| e^{{-|\partial_y} I_t|}.
]

The implementation uses a ResNet-50 encoder and 5-layer decoder with skip connections for the depth network, a ResNet-18 pose network, and a $3\times 3$ PatchGAN discriminator with $4\times 4$ kernels. Training was performed in PyTorch on four RTX2080Ti GPUs for $20$ epochs, with input resolutions $576\times 320$ for RobotCar and $768\times 384$ for nuScenes [2108.03830].

On RobotCar-Night, the reported metrics for DepthDark are Abs Rel $0.1205$, Sq Rel $0.5204$, RMSE $2.9015$, RMSE log $0.1633$, $\delta_1=0.8794$, $\delta_2=0.9688$, and $\delta_3=0.9896$. On nuScenes-Night, the reported metrics are Abs Rel $0.3150$, Sq Rel $3.7926$, RMSE $9.6408$, RMSE log $0.4026$, $\delta_1=0.5081$, $\delta_2=0.7776$, and $\delta_3=0.8959$. The paper also reports that removing the coordinate input $I_p$ in PBR worsens Sq Rel by $11.6\%$ on RobotCar and $15.1\%$ on nuScenes [2108.03830].

4. Foundation-model adaptation for low-light monocular depth

A later usage, “DepthDark: Robust Monocular Depth Estimation for Low-Light Environments” [2507.18243], repositions DepthDark as a foundation-model adaptation method. The backbone uses a pre-trained DINOv2 vision transformer as encoder and the DPT decoder, with initialization from Depth Anything V2. Instead of full fine-tuning, the method freezes the encoder and decoder and trains an additional low-light module with approximately $9$ K parameters.

A central component is Low-Light Dataset Generation (LLDG), which synthesizes nighttime data from daytime RGB-D pairs. Its Flare Simulation Module samples real flare patterns from Flare7K, randomly chooses flare centers
[
P_i = z_i\cdot K_I^{-1}\cdot [u_i\ v_i\ 1]^T,\qquad 0< z_i \le 20\,\mathrm{m},
]
and draws
[
s_b \sim \mathrm{Uniform}(0.4,1.0),\quad
s_F \sim \exp(U(\log s_F^{\min}, \log s_F^{\max})),
]
[
F \sim \exp(U(\log F^{\min}, \log F^{{\max})),\quad}
N_F = \max(\lfloor F/s_F+0.5\rfloor,1),
]
[
g_F \sim \mathrm{Uniform}(1.8,2.2).
]
The flare-augmented image is
[
I^F = (s_b\cdot I)^{g_F} + \sum_{i=1}^{N_F} [ss(L_S; s_F, P_i)]^{g_F}.
]

The Noise Simulation Module then injects sensor-level noise under a shot-read model:
[
I^{FN} = I^F + N,\qquad
N = K\cdot N_p + N_{\mathrm{read}} + N_r + N_q,
]
with $N_p \sim \mathrm{Poisson}(I^F)$, row noise $N_r \sim \mathcal N(0,\lambda_{\mathrm{row}})$, and quantization noise $N_q \sim \mathrm{Uniform}(-\tfrac12,\tfrac12)$ of variance $\lambda_{\mathrm{quant}}$ [2507.18243].

LLDG produces $74$ K paired images $(I^{{FN},\mathrm{depth})$} from Hypersim and Virtual KITTI, with flare count $N_F \in [1,20]$, brightness scale $s_b \in [0.4,1.0]$, gamma $g_F \in [1.8,2.2]$, and noise parameters chosen according to ELD guidelines. The method states that this yields distributions close to real nuScenes-Night and RobotCar-Night scenes [2507.18243].

The low-light parameter-efficient fine-tuning strategy adds illumination guidance and multiscale feature fusion. A single-channel illumination map is defined as
[
I_g^{FN} = \mathrm{mean}c(I^{FN}),
]
which is concatenated with the RGB image to form a four-channel tensor $I_A^{FN}$. Three parallel convolutional branches compute
[
E_1 = \mathrm{Conv}{1\times 1}(I_A^{FN}),\quad
E_2 = \mathrm{Conv}{3\times 3}(I_A^{FN}),\quad
E_3 = \mathrm{Conv}{5\times 5}(I_A^{FN}),
]
attention weights are obtained by
[
\alpha_i = \mathrm{Softmax}(W_i\cdot E_i + b_i),\qquad i=1,2,3,
]
and the fused feature is
[
E_{\mathrm{fused}} = \sum_{i=1}³ \alpha_i \odot E_i.
]
After a $1\times 1$ projection, the adapted feature is forwarded through the frozen PatchEmbed, transformer, and DPT decoder [2507.18243].

Training is performed in PyTorch on a single NVIDIA RTX 3090 at input resolution $518\times 518$, batch size $8$, using AdamW with learning rate $2\times 10^{-5}$ and weight decay $1\times 10^{-6}$. The reported training duration is approximately $4$ hours for $5$ epochs on $74$ K synthetic pairs [2507.18243].

On nuScenes-Night, the reported DepthDark metrics are Abs Rel $0.210$, Sq Rel $1.910$, RMSE $7.764$, RMSE log $0.260$, $\delta_1=0.630$, $\delta_2=0.914$, and $\delta_3=0.976$. On RobotCar-Night, the reported metrics are Abs Rel $0.157$, Sq Rel $1.063$, RMSE $4.284$, RMSE log $0.202$, $\delta_1=0.760$, $\delta_2=0.941$, and $\delta_3=0.985$ [2507.18243].

5. Dark-channel-assisted depth-from-defocus

In “Dark Channel-Assisted Depth-from-Defocus from a Single Image” [2506.06643], the label denotes a single-image depth-from-defocus method. The forward model is a spatially variant blur process:
[
I_{df}(x) = \int_\Omega G(x-y; r(d(y)))\,I(y)\,dy + \eta(x),
]
or in discrete form,
[
I_{df} = H(d)\star I + \eta.
]
The blur radius is linked to scene depth by
[
r(d) = \frac{1}{\sqrt{2}\,p_x}\cdot\frac{A f}{D_{fp}-f}\cdot\frac{|d-D_{fp}|}{d},
]
where $f$ is focal length, $A=f/F_n$ is aperture diameter, $p_x$ is sensor pixel pitch, and $D_{fp}$ is in-focus distance [2506.06643].

The method uses the dark channel prior
[
DC_I(p) = \min_{c\in{R,G,B}}\Bigl(\min_{q\in\Omega(p)} I_c(q)\Bigr),
]
with neighborhood size $N=15$. The paper argues that well-focused regions preserve local intensity contrasts and yield larger variance in the dark channel, whereas heavily blurred areas reduce that variance. On this basis it defines the Local Defocus and Dark-Channel Variation map:
[
\mathrm{LDDCV}(i) =
\left[
\max_{q\in N(i)} |I_{df}(i)-I_{df}(q)|,\;
\max_{q\in N(i)} |DC_{I_{df}}(i)-DC_{I_{df}}(q)|
\right],
]
where $N(i)$ is the $3\times 3$ neighborhood around pixel $i$. A validity mask is then defined by
[
M(i)=1[|\mathrm{LDDCV}(i)|>T],\qquad T=0.05.
]

The architecture combines several branches. A Dark-Channel Embedding network encodes $J_{df}=DC_{I_{df}}$ to a global feature vector $z\in\mathbb R^Q$. An LDDCV-Net and a mask-mediated sparse pooling network encode the two-channel LDDCV map and its mask into multi-scale features. These are fused with features from a pretrained ResNeXt101-32$\times$8d-wsl encoder through the Nested Feature Modulation & Fusion module, denoted Nest(FM)$^2$, which includes ARU, CFTB, MFEB, and HR$^2$B blocks. Multi-scale refinement then uses DSA-MSPF blocks before decoding to a continuous depth map [2506.06643].

Training includes adversarial supervision. The discriminator takes triples $[d,r(d),I_{df}]$ and distinguishes real from generated examples, yielding
[
L_{\mathrm{adv}} = \tfrac12 E_{I_{df},d_{gt}}[(D(d,r(d),I_{df})-1)^2].
]
The full objective is
[
L_{\mathrm{total}} = L_{\mathrm{spafid}} + 0.1L_{\mathrm{freq}} + 0.1L_{\mathrm{adv}},
]
where
[
L_{\mathrm{spafid}} = |d-d_{gt}|1,\qquad
L{\mathrm{freq}} = |\mathrm{DCT}(d)-\mathrm{DCT}(d_{gt})|_1.
]

The reported training set is the Eigen split of NYU-Depth v2, with $795$ train and $654$ test images. Synthetic defocus uses $f=9$ mm, $D_{fp}=0.7$ m, $F_n=2$, and $p_x=7.5\,\mu$m. Evaluation is performed zero-shot on $1{,}305$ real defocused images from EBD. The paper reports Absolute Rel Error $0.042$, Sq Rel Error $0.019$, RMSE $0.240$ m, log RMSE $0.032$, $\delta_1=0.975$, $\delta_2=0.995$, and $\delta_3=0.999$ [2506.06643].

6. All-day depth completion and closely related dark-scene geometry

The “All-day Depth Completion” line uses “DepthDark” as an end-to-end depth-completion system rather than a monocular estimator [2405.17315]. Its first stage, SpaDe, maps sparse projected LiDAR depth $S$ to a coarse dense depth and log-uncertainty:
[
D_c=f_\theta(S),\qquad U=g_\theta(S).
]
SpaDe is trained on synthetic data using a reconstruction term and a Gaussian heteroscedastic log-variance term:
[

\mathcal L_{\mathrm{SpaDe}}

|D_c-d^*|_2²
+
\frac12\sum_{i,j}
\left[
\exp(-U_{i,j})(D_{c,i,j}-d^{*{i,j})^2+U{i,j}}
\right].
]

The second stage, uncertainty-driven residual learning, refines the coarse estimate:
[
D_f = D_c + R_\phi(I,U).
]
An alternative fusion form uses an uncertainty gate
[
\lambda(U_{i,j}) = \frac{1}{1+\exp(\alpha(U_{i,j}-\beta))},
]
with typical values $\alpha=0.8$ and $\beta=0$, to combine $D_c$ and a learned dense prediction. The refinement loss is
[
\mathcal L_{\mathrm{URL}}=|D_f-d^*|_p + w_{\mathrm{sm}}\mathcal L_{\mathrm{sm}},
]
with edge-aware smoothness weighted by image gradients. SpaDe contains approximately $3$ M parameters and runs in approximately $8$ ms on a 3090 GPU at resolution $544\times 1600$. The full system is reported to run at approximately $30$ fps on a single 3090 [2405.17315].

On nuScenes, sample results with the MSG-CHN backbone show MAE improvement from $1674.46$ to $1332.28$ on daytime, from $2068.05$ to $1521.64$ on nighttime, and from $1713.83$ to $1343.23$ on all-day evaluation. Over three backbones, the URL scheme yields reported MAE gains of $11.65\%$ for all-day, $11.23\%$ for daytime, and $13.12\%$ for nighttime. The paper also states that SpaDe can be used in a plug-and-play fashion and allows a $25\%$ improvement when augmenting existing methods to preprocess sparse depth [2405.17315].

Two adjacent works extend the same research agenda without using the exact title “DepthDark.” “Dark3R: Learning Structure from Motion in the Dark” adapts MASt3R to raw images with SNR below $-4$ dB using teacher–student distillation on noisy–clean raw image pairs, requiring no 3D supervision. Its student is fine-tuned with LoRA adapters of rank $4$, and the reported training set includes approximately $42{,}000$ multi-view raw images with ground-truth 3D annotations [2603.05330]. “DarkVGGT: Seeing Through Darkness Using Thermal Geometry without Daylight Tax” introduces physics-inspired thermal factorization and geometry-shared thermal routing for RGB-T feed-forward geometry. On the reported average over three benchmarks, DarkVGGT improves AbsRel from $0.2010$ to $0.1373$, $\delta<1.25$ from $0.6901$ to $0.8243$, and RMSE_log from $0.2906$ to $0.2232$, while preserving well-lit RGB-only performance with $\delta<1.25$ drop below $1.5\%$ relative to VGGT [2606.11326].

A plausible implication is that the DepthDark label has progressively broadened from illumination compensation and nighttime monocular depth estimation to a wider program of dark-scene geometry that includes sparse-depth priors, raw-sensor processing, thermal cues, and foundation-model adaptation.

7. Shared principles, limitations, and conceptual distinctions

Across these variants, several design patterns recur. One is the replacement of explicit physical calibration with low-dimensional nuisance estimates or learned guidance. In the deep-sea pipeline, all unknown physics is folded into a static additive backscatter map and a dynamic multiplicative factor image estimated by medians rather than by modeling lamp positions, water parameters, or camera-light geometry [2110.00480]. In nighttime monocular depth estimation, low-light failure modes are handled by regularization, shared enhancement mappings, and dynamic masking rather than by assuming ideal photometric consistency [2108.03830]. In the foundation-model variant, low-light robustness is obtained through synthetic flare and noise generation and a small low-light adapter, leaving the bulk of the pretrained model frozen [2507.18243].

Another shared principle is modality compensation. The all-day depth-completion system uses LiDAR because sparse depth is illumination-invariant and in metric scale, while the image is dense but illumination-sensitive and scale-ambiguous [2405.17315]. DarkVGGT uses thermal sensing because long-wave infrared preserves structural edges and temperature patterns even in pitch-black conditions, while reflective residuals are isolated to reduce geometric ambiguity [2606.11326]. Dark3R moves to raw linear sensor data because conventional feature- and learning-based methods break down in the SNR regime below $-4$ dB [2603.05330].

The variants also differ sharply in supervision regime and task definition. The deep-sea method is a parameter-free preprocessing stage for albedo restoration and visual mapping, not a supervised depth predictor [2110.00480]. The 2021 nighttime monocular framework is self-supervised and optimized through view-synthesis losses regularized by adversarial distribution matching [2108.03830]. The 2025 low-light foundation-model version is trained on synthetic paired RGB-D data and uses PEFT [2507.18243]. The dark-channel variant addresses single-image defocus cues rather than nighttime imaging per se [2506.06643]. The all-day method assumes synchronized camera–LiDAR calibration and therefore belongs to depth completion rather than pure monocular estimation [2405.17315].

The limitations reported in the literature are likewise domain-specific. The self-supervised nighttime method notes that extremely dark or saturated scenes may still yield only coarse estimates [2108.03830]. The dark-channel depth-from-defocus method identifies homogeneous color regions and non-Lambertian surfaces as challenging, and assumes known camera parameters for blur synthesis [2506.06643]. DarkVGGT requires moderately aligned RGB-T pairs, and its additional hyperparameters may require calibration per sensor [2606.11326]. Dark3R depends on raw-image processing and noise-model calibration, and its reported robustness relies on noisy–clean raw training pairs, captured or synthesized [2603.05330].

Taken together, the literature presents DepthDark not as a single algorithmic object but as a technically diverse set of responses to the same underlying problem: depth, albedo, correspondence, and mapping become ill-posed when visible-light image formation is dominated by darkness, low SNR, defocus, flare, backscatter, or moving artificial illumination.