SeHDR: HDR Imaging and Reconstruction

• SeHDR comprises a family of context-dependent HDR formulations that enable single-exposure reconstruction, sensor-enhanced recovery, and HDR-based segmentation across varied imaging setups.
• It leverages techniques like latent exposure diffusion, spatially varying convolutions, and event-assisted fusion to restore radiance and mitigate saturation issues.
• SeHDR advances practical HDR applications by facilitating novel view synthesis with 3D Gaussian splatting and enhancing dynamic range restoration in challenging scenes.

Searching arXiv for papers and representative uses of “SeHDR” across HDR imaging, segmentation, and novel view synthesis. SeHDR denotes a family of high-dynamic-range formulations rather than a single universally fixed method. Across the cited literature, the label is used for HDR-based segmentation on radiance maps, single-exposure HDR reconstruction from one conventional LDR image, single-shot HDR from spatially varying exposures, sensor-enhanced or event-assisted HDR for dynamic scenes, and, as a proper method name, an HDR novel-view-synthesis framework based on 3D Gaussian Splatting (Dev et al., 2018, Wang et al., 2024, Guo et al., 2024, Li et al., 23 Sep 2025).

1. Terminological scope

In the cited research record, SeHDR is best understood as a context-dependent abbreviation centered on the use of HDR information for inference or reconstruction. The same label can refer to HDR as the representation being segmented, HDR as the target reconstructed from a single observation, HDR as the output of a sensor-enhanced dynamic-scene pipeline, or HDR as a radiance field inferred from single-exposure multi-view imagery.

Usage in the cited literature	Core formulation	Representative paper
HDR-based segmentation	Segment directly on HDR radiance maps	"High-Dynamic-Range Imaging for Cloud Segmentation" (Dev et al., 2018)
Single-exposure HDR	Recover HDR from one 8-bit LDR image	"LEDiff: Latent Exposure Diffusion for HDR Generation" (Wang et al., 2024)
Single-shot SVE HDR	Recover HDR from one raw capture with spatially varying exposures	"Single-Shot High Dynamic Range Imaging with Spatially Varying Exposures Considering Hue Distortion" (Go et al., 2019)
Selective exposure HDR	Choose a minimal exposure set for stack-based HDR acquisition	"Improved Image Selection for Stack-Based HDR Imaging" (Beek, 2018)
Sensor-enhanced / event HDR	Use events to align or reconstruct HDR in dynamic scenes	"Event-assisted 12-stop HDR Imaging of Dynamic Scene" (Guo et al., 2024)
Stereo / multi-view HDR	Build stereo HDR datasets or infer HDR 3D scene representations	"SeHDR: Single-Exposure HDR Novel View Synthesis via 3D Gaussian Bracketing" (Li et al., 23 Sep 2025)

A related acronym, SHDR, is used in the cited dataset literature for stereoscopic 3D HDR video and multi-exposure stereo resources rather than for single-exposure reconstruction per se (Banitalebi-Dehkordi, 2018, Choudhary et al., 2022). This suggests that SeHDR operates as a cluster of nearby concepts whose exact meaning is fixed by task, sensor model, and output space.

2. Image-formation models and reconstruction paradigms

A common starting point is the HDR imaging equation

$Z = f(E\Delta t),$

where $E$ denotes scene irradiance or radiance, $\Delta t$ the exposure time, $f(\cdot)$ a camera response function, and $Z$ the observed LDR intensity. In single-exposure HDR reconstruction, the task is to infer a plausible HDR image when only one $Z$ is observed; in single-shot SVE HDR, multiple effective exposures are encoded spatially inside one raw frame; in selective exposure HDR, preview data are used to decide which bracketed exposures should actually be captured (Wang et al., 2024, Zheng et al., 2021, Xu et al., 2021, Beek, 2018).

Within the single-exposure line, LEDiff repurposes a pretrained latent diffusion model to synthesize a latent exposure bracket $\{C_{-}, C_0, C_{+}\}$, fuse these latent codes with per-pixel, per-channel softmax-normalized weights, and decode the merged latent to a linear HDR image in log-radiance space. The method explicitly separates content priors from HDR radiance mapping, and it functions both as a generative HDR model and as an LDR-to-HDR converter (Wang et al., 2024). A different hybrid formulation restores saturated LDR regions by generating one darker synthetic image and one brighter synthetic image via camera-response-based intensity mapping functions, refining both with an Exposedness Aware Saturation Restoration Network, and then combining them through HDR synthesis or multi-scale exposure fusion (Zheng et al., 2021).

Single-shot SVE formulations move the multi-exposure problem into sensor readout. One strand treats hue preservation as the central issue: a single Bayer SVE image is separated into low- and high-exposure sub-images, the maximally saturated colors are estimated, and the fused result is corrected on the constant hue plane so that hue is borrowed from unsaturated measurements whenever possible (Go et al., 2019). Another strand compensates for the practical limitation that dual-ISO or dual-exposure sensors usually provide only two base exposures. It segments the scene in the joint luminance space of low- and high-exposure observations, rescales luminance per scene region to middle gray, generates $2S$ virtual exposures, and fuses them with a standard multi-exposure fusion operator (Go et al., 2019).

A more fully integrated single-shot formulation performs joint demosaicing and HDR reconstruction directly from an SVE Bayer radiance image. In that setting, the exposure pattern is row-dependent,

$$\Delta t_{ij}= \begin{cases} \tau_S, & i \bmod 4 = 1 \text{ or } 2\ \tau_L, & i \bmod 4 = 3 \text{ or } 0, \end{cases}$$

and the network uses spatially varying convolution to respect the Bayer-plus-exposure pattern while an exposure-guidance mask suppresses over- and under-exposed pixels during feature propagation (Xu et al., 2021).

A distinct acquisition-oriented usage treats SeHDR as selective exposure HDR. Here the problem is not to reconstruct HDR from one observation, but to choose a minimal set of exposures that fully covers scene irradiance while respecting SNR and saturation constraints. The preview stack is converted into a binary coverage matrix and solved as a set covering problem with the consecutive ones property, yielding polynomial-time exposure selection for stack-based HDR capture (Beek, 2018).

3. SeHDR as a prior inside dynamic multi-exposure HDR

SeHDR also appears as an assisting prior inside multi-exposure HDR pipelines. In this role, a ghost-free single-image HDR estimate is not the final product but a structural guide for multi-exposure fusion in dynamic scenes, where motion and saturation jointly destabilize alignment (Li et al., 2024, Li et al., 2023).

The clearest example is the SHDR-ESI / SHDR-A-MHDR framework. Its single-image branch first computes

$$\bm H_i = \frac{\bm L_i^\gamma}{t_i}, \qquad \bm X_i=[\bm L_i,\bm H_i],$$

with $E$0, and then reconstructs HDR from the reference image $E$1 assisted by an Enhanced Stop Image derived in IPT space. The ESI is built from

$E$2

followed by thresholding to emphasize strong chromatic boundaries. This ESI is processed jointly with the reference feature by a Self-Representation Module and a Mutual-Representation Module, which use cross-attention and dynamic modulation to recover detail in oversaturated regions (Li et al., 2024).

The resulting ghost-free SHDR feature stream then guides the multi-exposure branch. A Feature Interaction Fusion Module produces per-pixel, per-channel weights for combining reference and non-reference features, and a Ghost Suppression Module uses cross-attention between the SHDR feature and the fused multi-exposure feature to suppress inconsistent content. The method achieves the best PSNR-$E$3, PSNR-L, SSIM-L, and HDR-VDP-2 on Kalantari’s and Hu’s HDR datasets, and also the best BTMQI, MEF-SSIMd, and UDQM on Sen’s and Tursun’s datasets, indicating that a strong SeHDR prior can materially improve multi-exposure deghosting (Li et al., 2024).

A related but purely multi-exposure formulation is the scale-aware two-stage HDR framework. It linearizes each input with

$E$4

treats the middle exposure as reference, aligns non-reference features by a Spatial Correct Module using global context and deformable convolution, and then performs HDR synthesis in two stages. The first stage yields a preliminary ghost-light result; the second conflates that result with aligned features to remove residual artifacts. This architecture is explicitly designed for large-scale motion and saturation without explicit optical-flow alignment (Li et al., 2023).

4. HDR-based segmentation and scene analysis

In computer vision rather than image synthesis, SeHDR can denote segmentation performed directly on HDR data. The canonical example in the cited material is HDRCloudSeg for cloud segmentation in ground-based sky imagery (Dev et al., 2018).

The underlying imaging problem is radiometric. Whole sky imagers observe scenes in which the circumsolar region is extremely bright while the horizon and distant clouds are comparatively dark. Single LDR exposures therefore produce circumsolar saturation and horizon under-exposure, breaking color-ratio features commonly used for cloud segmentation. HDRCloudSeg addresses this by capturing three exposures at $E$5, $E$6, and $E$7, reconstructing an HDR radiance map with the Debevec–Malik algorithm, and segmenting on HDR-derived features rather than on a tone-mapped proxy (Dev et al., 2018).

The method computes the normalized red-blue ratio

$E$8

uses fuzzy C-means to estimate cloudness, converts high-confidence memberships into seeds with a threshold $E$9, and then solves a Boykov–Jolly graph-cut energy of the form

$\Delta t$0

The output is a binary sky/cloud mask over the entire cropped HDR scene (Dev et al., 2018).

Evaluation is performed on the SHWIMSEG dataset, which contains 52 HDR sets, each comprising three bracketed LDR images, a fused HDR radiance map, a tone-mapped version, and a manually annotated binary segmentation mask. On this benchmark, HDRCloudSeg operating on HDR radiance maps reports Precision $\Delta t$1, Recall $\Delta t$2, F-score $\Delta t$3, and Error $\Delta t$4, outperforming the same method on medium-exposure LDR and on tone-mapped HDR as well as four LDR baselines. The radiance maps also exhibit 24× fewer saturated pixels than high-exposure LDR and 4× fewer than medium-exposure LDR, which is central to its circumsolar performance (Dev et al., 2018).

This use of SeHDR is methodologically distinct from single-image HDR reconstruction. The segmentation target is not HDR itself but a binary labeling problem solved more reliably once radiance information has been restored.

5. Sensor-enhanced, stereo, and multi-view expansions

A major extension of SeHDR replaces the single conventional frame with additional sensing modalities. In event-assisted 12-stop HDR imaging, an RGB camera is paired with a Prophesee EVK4-HD event camera via a beam splitter and hardware synchronization. Five LDR frames at $\Delta t$5, $\Delta t$6, $\Delta t$7, $\Delta t$8, and $\Delta t$9 EV are aligned with event-conditioned flow, pre-reconstructed to reduce blur and noise, and fused by a diffusion-based HDR module conditioned through ControlNet. On the synthetic ESHDR benchmark, the resulting system reports PSNR $f(\cdot)$0, SSIM $f(\cdot)$1, and LPIPS $f(\cdot)$2, while on 16 real scenes it achieves MUSIQ $f(\cdot)$3, MANIQA $f(\cdot)$4, and HyperIQA $f(\cdot)$5 (Guo et al., 2024).

A self-supervised event-based variant, Self-EHDRI, addresses the absence of sharp HDR ground truth in highly dynamic scenes. Its E-BL2SH network predicts a sharp HDR frame sequence $f(\cdot)$6 from one blurry LDR image and concurrent events, while a decomposition/composition assistance branch enforces HDR–LDR, LDR–LDR, LDR–HDR, and HDR–HDR consistencies. The framework is trained on BL2SHD-S and BL2SHD-R, and on BL2SHD-Rsimple it reports PSNR-$f(\cdot)$7 $f(\cdot)$8, SSIM-$f(\cdot)$9 $Z$0, and HDR-VDP-2 $Z$1 without using sharp HDR ground truth during the target dynamic-scene regime (Xiaopeng et al., 2024).

Stereo and multi-view settings broaden SeHDR from 2D reconstruction to immersive imaging and scene representation. The SHDR video database provides 11 stereoscopic HDR videos, each about 10 seconds long, at $Z$2, 30 fps, and 18 F-stops, stored as floating-point linear RGB .hdr streams and as 12-bit 4:2:0 BT.709 .yuv streams. It is explicitly intended for stereo HDR quality assessment, coding, tone mapping, and depth-related studies (Banitalebi-Dehkordi, 2018). A later IIT Madras dataset contributes 39 stereo multi-exposure image scenes and 18 stereo multi-exposure video scenes, with $Z$3 image views, $Z$4 video views at 30 fps, 3–4 exposures for images, 3–6 exposures for videos, and ZED-derived depth maps, thereby supporting stereo HDR reconstruction, consistent tone mapping, and HDR-aware depth estimation (Choudhary et al., 2022).

As a proper method name, SeHDR has also been introduced for HDR novel view synthesis from single-exposure multi-view LDR images. In that formulation, base 3D Gaussians are learned in a linear color space, bracketed 3D Gaussians with identical geometry but varying linear colors are synthesized according to

$Z$5

and a Differentiable Neural Exposure Fusion module fuses them into HDR Gaussians. On the synthetic HDR-NeRF benchmark, SeHDR without HDR ground truth reports PSNR $Z$6, SSIM $Z$7, and LPIPS $Z$8, substantially improving over single-exposure adaptations of HDR-NeRF and HDR-GS (Li et al., 23 Sep 2025).

6. Evaluation practices, limitations, and research directions

Evaluation protocols vary with the SeHDR sense being used. Segmentation-oriented work uses Precision, Recall, F-score, Error, and ROC quantities such as TPR and FPR (Dev et al., 2018). Single-image and event-assisted HDR reconstruction emphasize HDR-VDP-2 or HDR-VDP3, PU2-PSNR, PU2-SSIM, PU21-PIQE, FID on tone-mapped outputs, LPIPS, DISTS, MANIQA, MUSIQ, HyperIQA, NIQE, and CLIP-IQA+ (Zheng et al., 2021, Wang et al., 2024, Guo et al., 2024, Xiaopeng et al., 2024). Exposure-fusion-style assessments additionally use MEF-SSIM, MEF-SSIMd, BTMQI, UDQM, and TMQI (Go et al., 2019, Li et al., 2024).

The cited papers also expose a set of recurrent limitations. CRF dependence and camera-specific calibration remain explicit constraints in hybrid single-image saturation restoration; inaccurate CRFs degrade synthetic exposure quality (Zheng et al., 2021). LEDiff inherits Stable Diffusion biases and failure modes, and its training setup does not explicitly address JPEG artifacts, sensor noise, or heavy compression typical of internet photographs (Wang et al., 2024). HDRCloudSeg remains challenged by extremely thin or translucent clouds near the sun, and by hazy conditions that alter the discriminative power of $Z$9 (Dev et al., 2018). Event-assisted 12-stop HDR and Self-EHDRI both identify domain gaps between simulated and real events, as well as residual difficulty under extreme motion, sparse events, or blur at high-EV exposures (Guo et al., 2024, Xiaopeng et al., 2024).

Multi-view and stereo formulations introduce further assumptions. The SHDR database offers only 11 sequences and does not include distributed depth maps or explicit HDR10-style metadata (Banitalebi-Dehkordi, 2018). The IIT Madras stereo dataset provides pseudo ground-truth depth from the ZED pipeline rather than explicit disparity ground truth, and its multi-exposure captures are sequential, so motion between exposures is intrinsic to the data (Choudhary et al., 2022). The 3D Gaussian SeHDR method assumes static scenes and relies on an approximate CRF chosen from DoRF; its synthetic bracketing remains an extrapolation from single-exposure observations (Li et al., 23 Sep 2025).

The research directions identified in the cited literature are correspondingly broad: stereo-aware tone mapping and 3D-HDR quality models for immersive displays, faster diffusion or distilled HDR generators, hybrid RGB-event sensors and temporally consistent HDR video, CRF-robust or CRF-agnostic single-image HDR, and physically grounded multi-view HDR representations that do not require explicit exposure bracketing (Banitalebi-Dehkordi, 2018, Choudhary et al., 2022, Wang et al., 2024, Guo et al., 2024, Li et al., 23 Sep 2025). Taken together, these works position SeHDR not as a single algorithmic family, but as a technically diverse research domain organized around the recovery, use, or exploitation of HDR information under acquisition constraints that range from single conventional frames to event streams and multi-view radiance fields.