LENVIZ: A High-Resolution Low-Exposure Night Vision Benchmark Dataset

A large-scale multi-exposure benchmark dataset of over 234K frames for low-light image enhancement and multi-exposure fusion (WACV 2026).

LENVIZ dataset overview: multi-exposure capture brackets, hardware illuminance estimation, and curated ground truth for low-light enhancement.

Overview

Capturing imagery in low-illumination environments introduces physical degradations: sensor noise, low signal-to-noise ratio (SNR), limited dynamic range, exposure clipping, and loss of color fidelity. Deep learning methods can address these issues, but their performance depends on the quality, scale, and diversity of paired training data.

To support research in this area, LENVIZ (Low Exposure Night Vision Dataset) provides a benchmark dataset comprising over 234,000 frames across 24,000 real-world scenes in resolutions up to 4K. Published at the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026), LENVIZ addresses gaps in existing benchmarks by providing multi-exposure capture brackets, multi-sensor coverage, ISP-tuned representations, and expert-retouched ground truth.

Dataset characteristics and contributions

1. Scale, multi-exposure brackets, and resolution

  • 234,688 total frames: Spans 24,000 indoor and outdoor scenes for training single-exposure and multi-exposure enhancement networks.
  • High resolution (up to 4K): Captured at native resolutions of $4080 \times 3072$ and $3264 \times 2448$, preserving fine textural details that are lost in downsampled datasets.
  • 9 exposure brackets and long exposure: Each scene includes up to 9 intermediate multi-exposure frames alongside a dedicated long-exposure ground truth shot, supporting single-frame enhancement and multi-exposure fusion (MEF).

2. Multi-sensor diversity

LENVIZ incorporates captures across three physical camera modules to evaluate cross-sensor generalization:

  • Front-facing module (S5K4H7YX03-FGX9): Native resolution of $3264 \times 2448$, intended for close-range captures (arm’s length) with a high concentration of facial subjects.
  • Rear sensor modules (S5KJN1SQ03 and S5KJNS): Native resolutions of $4080 \times 3072$, used for medium- and long-range environmental captures covering landscapes, architecture, and varying depth fields under low illumination.

3. ISP-tuned only pipeline

Most existing datasets provide either raw sensor data (which require proprietary camera decoders and are limited in deployment) or heavily processed JPEGs (which undergo aggressive vendor smoothing, noise reduction, and beautification filters).

LENVIZ introduces an ISP-tuned only pipeline:

  • Captured frames undergo sensor Image Signal Processing (ISP) without default camera post-processing filters.
  • This preserves the authentic noise profiles, dynamic range limits, and sensor artifacts encountered by production algorithms, bridging the gap between raw sensor data and deployment.

4. Controlled exposure and ambient illuminance formulation

Capture parameters (ISO sensitivity, exposure time, aperture) were systematically regulated, with scene illuminance ($L’$) quantitatively estimated from hardware parameters:

\[L' = \beta_{\text{illuminance}} \times f^2 \times \frac{10^{11}}{t_{\text{exp}} \times \text{ISO}}\]

Long-exposure thresholds ($\gamma_0$ through $\gamma_4$) were calibrated across illuminance brackets ranging from extreme darkness ($<1$ lux) to low ambient light ($\ge 15$ lux) to capture signal while minimizing motion blur.

5. Expert human-curated ground truth

To address limitations in synthetic pairs or noisy long-exposure captures, LENVIZ includes 13,067 scenes with human-curated ground truth:

  • Standardized retouching protocol: A team of 7 professional photographers and retouchers used the long-exposure shot as a clean structural base, restored clipped highlights using multi-exposure brackets, and normalized color fidelity against mid-exposure frames.
  • Quality verification: Every curated ground truth image was reviewed by a panel of three camera scientists across strict criteria: sharpness, noise level, dynamic range, and absence of haloing, color fringing, or ghosting. Only images receiving unanimous approval were accepted into the benchmark.

6. Demographic coverage and scene diversity

  • Skin-tone diversity: To evaluate fair facial feature restoration, the dataset incorporates 14 realistic mannequins covering a diverse range of skin tones, poses, and accessories (glasses, facial hair). Mannequins ensured motion-free stabilization during multi-exposure bursts.
  • Face concentration: Over 70% of scenes contain one or more faces.
  • Contextual metadata: Captured scenes were processed with automated label detection to provide bounding boxes, facial attributes, and 230 object classes across 27 scene categories.
  • Environments: Features indoor environments (mixed incandescent and fluorescent lighting, dark corners, high contrast) and outdoor scenarios (urban nightscapes, streetlamps, directional light, low ambient illumination).

Evaluation benchmark

LENVIZ includes a test benchmark of 1,468 frames across 203 unique scenes, divided into two complementary partitions:

  1. Reference partition (610 frames, 60 scenes): Paired with human-curated ground truth and long-exposure references for full-reference metric evaluation.
  2. No-reference partition (858 frames, 143 scenes): Incorporates handheld captures and live human subjects under natural motion to test model robustness and generalization.
  3. DxOMark test chart protocols: Includes test chart captures to diagnose color fidelity, texture preservation, and dynamic range recovery.

Benchmarking and empirical findings

The dataset was used to benchmark single-exposure and multi-exposure low-light restoration architectures:

  • Single-exposure baselines: LLFormer, ExpoMamba, ZeroDCE++
  • Multi-exposure baselines: MEFNet, HoLoCo, MobileMEF
  • Comparative training: Models were trained and evaluated against LOL and SICE.

Results

  • Perceptual and structural metrics: Models trained on LENVIZ achieved higher scores in LPIPS and SSIM, restoring natural textures and realistic illumination more consistently than models trained on prior datasets.
  • Cross-sensor generalization: Models trained on LENVIZ generalized effectively across unseen camera sensors and unfamiliar ISP configurations.
  • User study: A double-blind pairwise study (A/B testing) conducted with 238 human evaluators (sampled from a pool of 89,883 candidates) showed that evaluators preferred LENVIZ-trained models (e.g., 84.44% preference for ExpoMamba trained on LENVIZ over the same model trained on LOL), citing realistic brightness and natural color reproduction.

Paper and resources

Conference paper

  • Title: LENVIZ: A High-Resolution Low-Exposure Night Vision Benchmark Dataset
  • Authors: M. Aithal, R. G. VidalMata, M. Kartha, G. Chen, E. Adhikarla, L. N. Kirsten, Z. Fu, N. A. Madusudhana, J. Nasti
  • Venue: IEEE/CVF Winter Conference on Applications of Computer Vision (WACV 2026)
  • OpenAccess PDF: CVF OpenAccess Paper

Resources