SmokeSVD: Dual Approaches in Smoke Analysis

• SmokeSVD is an overloaded term describing two distinct applications: an industrial smoke detection and visualization system and a framework for monocular dynamic smoke reconstruction.
• The detection version uses heuristic computer vision techniques, including change detection and texture segmentation, to efficiently process fixed camera feeds and generate interactive visual summaries.
• The reconstruction version employs physics-guided diffusion, progressive novel-view synthesis, and differentiable advection under Navier–Stokes constraints to recover 3D smoke dynamics from a single view.

to=arxiv_search 天天中彩票是 ашәҟәы=json {"query":"SmokeSVD smoke reconstruction single view diffusion models arXiv", "max_results": 5, "sort_by": "submittedDate"} to=arxiv_search eqqarsியம் to=json {"query":"Industrial Smoke Detection and Visualization arXiv SmokeSVD", "max_results": 5, "sort_by": "submittedDate"} to=arxiv_search eqqarsியம் to=json {"query":"Self-Supervised Video Desmoking for Laparoscopic Surgery SelfSVD arXiv", "max_results": 5, "sort_by": "submittedDate"} SmokeSVD is an overloaded term in the smoke-analysis literature rather than a single universally standardized name. In one explicit usage, "SmokeSVD" abbreviates Smoke Detection and Visualization, a system for detecting daytime fugitive emissions from a static refinery camera and presenting them through timeline navigation, autonomous fast-forwarding, and animated summaries (Hsu et al., 2018). In a later and unrelated usage, "SmokeSVD" is the title of a framework for smoke reconstruction from a single view, combining physics-guided diffusion-based side-view synthesis, progressive novel-view refinement, and differentiable advection to recover dynamic volumetric smoke from a single RGB video (Li et al., 16 Jul 2025). The term is also often conflated with neighboring names such as SelfSVD, but those are distinct official designations rather than alternate spellings of SmokeSVD (Wu et al., 2024).

1. Nomenclature and scope

The literature contains two direct, official uses of the string "SmokeSVD," and several adjacent works that are sometimes informally confused with it. The resulting ambiguity is substantive because the underlying tasks differ: one concerns event detection and visualization from fixed cameras, while the other concerns 3D dynamic reconstruction from monocular video.

Usage	Official meaning	Source
SmokeSVD	Smoke Detection and Visualization	(Hsu et al., 2018)
SmokeSVD	Smoke reconstruction from a single view via diffusion and refinement	(Li et al., 16 Jul 2025)
SelfSVD	Self-Supervised Video Desmoking for Laparoscopic Surgery; not called SmokeSVD by its authors	(Wu et al., 2024)

This dual usage makes disambiguation essential. In industrial monitoring contexts, SmokeSVD refers to a detection-and-visualization workflow built around smoke masks and event summaries. In computational fluid reconstruction contexts, SmokeSVD refers to a physically constrained generative pipeline that estimates density, velocity, and source terms from a single video. A plausible implication is that "SmokeSVD" should not be treated as a stable task label without identifying the specific paper or application domain.

2. SmokeSVD as Smoke Detection and Visualization

In "Industrial Smoke Detection and Visualization," SmokeSVD is an end-to-end system for detecting daytime fugitive emissions from a static camera at a coke refinery and exposing the results through three visualization features: an interactive timeline, an autonomous fast-forwarding mode, and automatically generated animated smoke images (Hsu et al., 2018). The detector processes daytime frames inside a fixed ROI of $496\times 528$ pixels, downsamples by a factor of 4 for speed, computes a per-frame smoke mask $M_t$, and converts that mask into a scalar response equal to the number of smoke pixels.

The detection pipeline has five stages: preprocessing, change detection, texture segmentation, region filtering, and event detection. Preprocessing builds a temporal-median background model over the previous 60 frames,

$$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$

and uses the illumination-compensated subtraction

$\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$

Change detection combines a high-frequency branch based on DoG filtering and local entropy with an intensity branch based on CLAHE and frame/background differencing. Texture segmentation then applies Laws’ texture-energy filters, PCA preserving 98% variance, and accelerated k-means++ clustering into textons or regions. Region filtering removes candidates using shape, color, size, change strength, and a shadow discriminator based on KDE over subtracted intensities. Event detection groups smoke-positive frames into segments and merges nearby segments into events.

The system is operationalized as a visual-analytics tool rather than only a detector. The timeline visualizes the detector’s time series as clickable spikes and segments. Fast-forwarding plays only segments predicted to contain smoke. Animated image generation compiles short event-centric summaries suitable for documentation and sharing. The paper reports processing time of about 30 minutes per day on a dual hex-core Intel Xeon X5670 workstation over 9,700 daytime frames, and presents day-level precision, recall, and F-score examples such as $0.7500/0.9474/0.8372$ on Nov 15 and $0.5400/0.9643/0.6923$ on Oct 05. False positives are dominated by steam and fast-moving shadows, while false negatives arise for low-opacity or small smoke. Nighttime detection is not supported.

The significance of this version of SmokeSVD lies in its coupling of heuristic CV with community-facing visualization. It is not a learned model and does not estimate physical smoke state; instead, it reduces manual review burden and turns detection outputs into navigable evidence for environmental monitoring and citizen-science workflows.

3. SmokeSVD as single-view dynamic smoke reconstruction

In "SmokeSVD: Smoke Reconstruction from A Single View via Progressive Novel View Synthesis and Refinement with Diffusion Models," SmokeSVD denotes a monocular reconstruction framework for dynamic smoke (Li et al., 16 Jul 2025). The problem is posed as severely ill-posed because many distinct 3D density fields can explain the same 2D projection. The framework addresses this with three major stages: physically guided side-view synthesis, progressive novel-view synthesis with iterative refinement, and final fine-scale reconstruction via differentiable advection under Navier–Stokes constraints.

The formulation uses smoke density $\rho(x,t)$, velocity $u(x,t)$, and source $s(x,t)$. The front-view input at time $t$ is $M_t$0, the synthesized side-view is $M_t$1, and novel views are $M_t$2. First, a diffusion-based side-view synthesizer, SvDiff, predicts side views frame by frame from the current front view and prior synthesized or rendered side views. Its training objective combines a noise-prediction term, image fidelity, spatial distribution consistency, and velocity guidance:

$M_t$3

The velocity guidance penalizes both incompressibility violation and excessive gradients:

$M_t$4

Second, from the front and side views, a coarse density field is reconstructed and rendered to increasingly larger novel angles. These rendered views are refined by NvRef and then fed back to improve the coarse density estimate. The method caps the number of views at 16 and stages them from near to mid to far angles. Third, with a higher-quality multi-view set, the framework reconstructs fine density and jointly estimates velocity and source via differentiable advection. The physical constraints include the advection-diffusion equation for density,

$M_t$5

the incompressible Navier–Stokes momentum equation,

$M_t$6

and incompressibility,

$M_t$7

This final stage is designed to produce physically coherent long-term dynamics and enable re-simulation.

The empirical evaluation uses ScalarFlow and a synthetic dataset. On ScalarFlow, the paper reports for SmokeSVD: Input RMSE $M_t$8, SSIM $M_t$9, PSNR $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$0, LPIPS $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$1, Side RMSE $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$2, STYLE $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$3, and runtime 15 minutes for 120 steps. GlobTrans attains stronger perceptual metrics but requires more than 30 hours. On the synthetic dataset, SmokeSVD reports RMSE $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$4, SSIM $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$5, PSNR $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$6, LPIPS $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$7, Side RMSE $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$8, and STYLE $$B_t(x,y)=\mathrm{median}\big(I_t(x,y),\ldots,I_{t-59}(x,y)\big),$$9. The method’s stated advantage is therefore not only reconstruction quality but the combination of multi-view consistency, physical plausibility, and substantially reduced runtime relative to optimization-heavy baselines.

4. Relation to adjacent smoke datasets and video methods

Several nearby works clarify what SmokeSVD is not. "Multiple Categories Of Visual Smoke Detection Database" introduces the Three-Categories Smoke Detection Database (TCSDD) rather than a database called SmokeSVD; its contribution is a 70k-scale, three-class industrial smoke image dataset aligned with operational states: smokeless, black smoke, and white smoke (Gong et al., 2022). "Video-based Smoky Vehicle Detection with A Coarse-to-Fine Framework" introduces LaSSoV and LaSSoV-video, plus the CoDeS pipeline for smoky vehicle detection, and explicitly does not use SmokeSVD as a dataset name (Peng et al., 2022).

In laparoscopic desmoking, "Self-Supervised Video Desmoking for Laparoscopic Surgery" introduces SelfSVD, not SmokeSVD (Wu et al., 2024). SelfSVD uses a pre-smoke frame both as unaligned supervision and as a reference input. Its core reconstruction loss is

$\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$0

supplemented by an $\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$1 regularization on masked reference features and an LSGAN objective. On the LSVD test set, SelfSVD* reports PSNR $\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$2 dB and SSIM $\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$3, while the lightweight SelfSVD-S runs at $\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$4 s/frame with 1.92 M parameters. The paper notes that some readers may colloquially refer to SelfSVD as "SmokeSVD," but this is not the authors’ terminology.

A later surgical work, "Rethinking Surgical Smoke: A Smoke-Type-Aware Laparoscopic Video Desmoking Method and Dataset," introduces STANet and the STSVD dataset rather than SmokeSVD (Liang et al., 2 Dec 2025). It defines two smoke types—Diffusion Smoke and Ambient Smoke—and reports on STSVD that STANet reaches PSNR $\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$5, SSIM $\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$6, and LPIPS $\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$7, outperforming SelfSVD on that benchmark. Collectively, these works show that SmokeSVD is not a universal umbrella acronym for smoke-related vision datasets or desmoking methods; the field uses task-specific names tied to industrial monitoring, vehicle surveillance, laparoscopic restoration, or dynamic fluid reconstruction.

5. Methodological contrasts and common misconceptions

The two official SmokeSVD usages differ not only in application but in ontological target. The 2018 SmokeSVD estimates a binary smoke mask $\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$8, a scalar per-frame response, and event segments from a fixed camera stream (Hsu et al., 2018). The 2025 SmokeSVD estimates volumetric density $\mathrm{bgSub}(I,B)=\frac{I-B}{\max(I+B,0.1)}.$9, velocity $0.7500/0.9474/0.8372$0, and source $0.7500/0.9474/0.8372$1 from a single RGB video, and is explicitly designed to support re-simulation and downstream 3D representations (Li et al., 16 Jul 2025). One therefore operates in a heuristic 2D event-detection regime; the other operates in a generative, physics-constrained 3D inverse-problem regime.

A common misconception is to read SmokeSVD as a generic label for smoke video processing. The neighboring literature contradicts that simplification. TCSDD is an image-level industrial classification dataset; LaSSoV targets smoky vehicle detection with a YOLO-based detector, smoke-vehicle matching, and a short-term 3DCNN; SelfSVD is a self-supervised surgical video restoration method based on pre-smoke supervision; STANet is a smoke-type-aware laparoscopic desmoking network (Gong et al., 2022). This suggests that SmokeSVD functions more as a paper-specific name than as a taxonomic category.

Another misconception is to conflate the 2025 SmokeSVD with standard novel-view synthesis for rigid objects. The paper explicitly positions its method against that analogy: smoke is semi-transparent, dynamic, and physically constrained, so cross-view consistency cannot rely on rigid-scene assumptions alone (Li et al., 16 Jul 2025). Likewise, conflating the 2018 SmokeSVD with modern deep smoke detectors would obscure its heuristic design; it intentionally avoids optical flow and deep training in favor of change detection, texture segmentation, and region filtering suited to an operational citizen-science tool (Hsu et al., 2018).

6. Limitations, impact, and future directions

The limitations of the 2018 SmokeSVD are explicit: it relies on heuristic thresholds, supports daytime only, and remains vulnerable to steam, shadows, low-opacity smoke, lighting variation, and static-camera assumptions (Hsu et al., 2018). The paper identifies future directions such as crowdsourced labels, learned classifiers, meteorological integration, and multi-sensor inputs. Its impact is primarily practical and socio-technical: it reduces manual review time, supports documentation, and helps community members focus on interpretation and storytelling rather than exhaustive frame-by-frame search.

The 2025 SmokeSVD faces a different class of limitations. Single-view constraints remain difficult for highly complex, fast turbulent flows; far-angle synthesized images may exhibit subtle artifacts; sensitivity to camera calibration and lighting mismatch can affect rendering-based supervision; and extension beyond smoke to participating media with stronger scattering would require extended physics and appearance models (Li et al., 16 Jul 2025). Even so, the framework represents a shift from monocular smoke observation to monocular smoke reconstruction with physically meaningful latent variables and re-simulatable dynamics.

Placed together, the two SmokeSVD lineages map a broader evolution in smoke-related computer vision. The earlier usage emphasizes detection, visualization, and human-in-the-loop environmental monitoring. The later usage emphasizes generative inversion, multi-view consistency, and physically guided 3D recovery. A plausible implication is that future work may continue to bridge these strands: operational monitoring systems may become more physically informed, while reconstruction systems may become more deployment-oriented, faster, and more robust to real-world sensing constraints.