Research map / Video / Video segmentation and tracking
Seeing Dark Videos via Self-Learned Bottleneck Neural Representation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a self-supervised, data-free low-light video enhancement method combining a hybrid neural representation, an entropy-based objective bottleneck, and a Chromatic Retinex decomposition.
Problem
Low-light video enhancement usually needs paired supervised data, which is costly to collect and creates domain gaps; existing self-supervised denoising and unsupervised enhancement methods struggle with intensive, hybrid noise and temporal flicker in dark videos.
Approach
Constrain self-supervised learning from both content and objective sides: a compact learned per-frame deep embedding feeds a video-level network (hybrid neural representation), an entropy constraint (Gaussian-mixture likelihood plus soft-histogram entropy and cross-entropy) suppresses noise and biased illumination, and a chromatic Retinex decomposition aligns reflectance temporally to enable noise
Key results
- Achieves NIQE 4.7616 and ILNIQE 31.1489 on the DRV evaluation set, the best among compared methods (Zero-DCE NIQE 5.1637).
- Best NIQMC of 4.8123, slightly above RUAS at 4.9217? No: RUAS reports 4.9217 higher, so the method is second on NIQMC at 4.8123.
- Ablation: removing the hybrid neural representation raises NIQE from 4.7616 to 8.5161.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Ours instead of Zero-DCE | low-light video enhancement · DRV evaluation set | NIQE: 5.1637 -> 4.7616 | Helped |
| use Ours instead of RUAS | low-light video enhancement · DRV evaluation set | ILNIQE: 33.0497 -> 31.1489 | Helped |
| use Ours instead of RUAS | low-light video enhancement · DRV evaluation set | NIQMC: 4.9217 -> 4.8123 | Hurt |
| use Ours instead of VBM4D + EnlightenGAN | cascaded denoising and enhancement · DRV evaluation set | NIQE: 4.8043 -> 4.7616 | Helped |
| use Ours instead of UDVD + RUAS | cascaded denoising and enhancement · DRV evaluation set | ILNIQE: 33.5089 -> 31.1489 | Helped |
| use Ours instead of UDVD + RUAS | cascaded denoising and enhancement · DRV evaluation set | NIQMC: 4.7444 -> 4.8123 | Helped |
| add hybrid neural representation (ablation) | Seeing Dark Videos via Self-Learned Bottleneck Neural Representation · DRV evaluation set | NIQE: 8.5161 -> 4.7616 | Helped |
| add Gaussian-mixture model objective bottleneck (ablation) | Seeing Dark Videos via Self-Learned Bottleneck Neural Representation · DRV evaluation set | NIQE: 5.1642 -> 4.7616 | Helped |
| add soft histogram constraint (ablation) | Seeing Dark Videos via Self-Learned Bottleneck Neural Representation · DRV evaluation set | NIQE: 5.6177 -> 4.7616 | Helped |
| use Chromatic Retinex instead of traditional Retinex (ablation) | Seeing Dark Videos via Self-Learned Bottleneck Neural Representation · DRV evaluation set | NIQE: 4.9132 -> 4.7616 | Helped |
| add channel-wise Gamma estimation (ablation) | Seeing Dark Videos via Self-Learned Bottleneck Neural Representation · DRV evaluation set | NIQE: 5.7324 -> 4.7616 | Helped |
Surprising findings
- Retraining unsupervised methods on the single test sequence from scratch gave no obvious gain, attributed to small training sample scale.
- Cascading unsupervised denoising (VBM4D, UDVD) with low-light enhancers did not guarantee better results than the proposed method and can over-smooth.
Limitations
- Relies on no-reference metrics (NIQE, ILNIQE, NIQMC) only, without reference-based or human evaluation.
- Evaluated on a single small dataset (DRV) with 100-120 frame clips.
- Unsupervised baselines could not be retrained effectively on single sequences, limiting fair comparison.
Experimental setup
Datasets: DRV
Benchmarks: DRV
Metrics: NIQE ILNIQE NIQMC
Baselines: Zero-DCE RUAS EnlightenGAN SMID SDSD VBM4D UDVD MF LIME
Builds on: NeRV HNeRV Noise2Noise Zero-DCE Retinex
Closest papers by meaning
- Learning Temporal Consistency for Low Light Video Enhancement From Single Images (CVPR 2021 · similarity 0.86)
- Self-Supervised Learning of Compressed Video Representations (ICLR 2021 · similarity 0.85)
- No More Shortcuts: Realizing the Potential of Temporal Self-Supervision (AAAI 2024 · similarity 0.85)
- Self-Supervised Learning for Color Spike Camera Reconstruction (CVPR 2025 · similarity 0.84)
- Learning predictable and robust neural representations by straightening image sequences (NeurIPS 2024 · similarity 0.83)
- Binarized Low-light Raw Video Enhancement (CVPR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.