Research map / Video / Video segmentation and tracking
Binarized Low-light Raw Video Enhancement
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.
Applies binary neural networks to low-light raw video enhancement using a distribution-aware binary convolution and a parameter-free spatial-temporal shift operation.
Problem
Low-light raw video enhancement models are computationally expensive and memory-heavy, hindering deployment on resource-limited devices; binarizing them raises issues of fusing temporal information and the accuracy gap of binary convolutions.
Approach
Introduces a cyclic temporal plus spatial shift operation that is parameter-free and binarization-friendly for temporal fusion, and a distribution-aware binary convolution that rescales binary conv outputs with a dynamic channel-attention factor computed from activation statistics (mean, mean of absolute value, std).
Key results
- On LLRVD, BRVE reaches 37.07 dB PSNR and 0.9581 SSIM, matching full-precision SMOID (37.07 dB) and beating EMVD-S (36.70 dB) with similar FLOPs.
- On SMOID at gain 0, BRVE achieves 40.05 dB PSNR with 0.3M params and 1.49 GFLOPs, improving over the BBCU binary baseline (39.92 dB) at 1.4% extra computation.
- Removing the shift operation drops LLRVD PSNR from 37.07 to 36.70 dB.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add distribution-aware convolution (DABC) instead of BBCU | BRVE on SMOID · SMOID Gain0 | PSNR: 39.92 -> 40.05 | Helped |
| add distribution-aware convolution (DABC) instead of BBCU | BRVE on SMOID · SMOID Gain15 | PSNR: 40.11 -> 40.25 | Helped |
| add distribution-aware convolution (DABC) instead of BBCU | BRVE on SMOID · SMOID Gain30 | PSNR: 40.48 -> 40.64 | Helped |
| add mean(|·|) statistics to DACA (ablation) | BRVE · LLRVD | PSNR: 36.95 -> 37.02 | Helped |
| add mean and standard deviation statistics to DACA (ablation) | BRVE · LLRVD | PSNR: 36.95 -> 37.01 | Helped |
| add the spatial-temporal shift operation (ablation) | BRVE · LLRVD | PSNR: 36.70 -> 37.07 | Helped |
| set sliding window stride to 2 (ablation) | BRVE · LLRVD | PSNR: 37.07 -> 36.95 | Hurt |
| set sliding window stride to 2 (ablation) | BRVE · LLRVD | FLOPs: 1.49 -> 0.89 | Helped |
| set sliding window stride to 3 (ablation) | BRVE · LLRVD | PSNR: 37.07 -> 36.86 | Hurt |
| set sliding window stride to 3 (ablation) | BRVE · LLRVD | ST-RRED: 0.0455 -> 0.0502 | Hurt |
Surprising findings
- Raw2Raw enhancement before ISP outperforms jointly learning ISP (Raw2RGB) and RGB-to-RGB enhancement, and H264 compression of low-light input severely degrades results.
Limitations
- Larger temporal stride reduces FLOPs but degrades PSNR and temporal consistency (ST-RRED rises)
- Evaluated only on two low-light raw video datasets
Experimental setup
Datasets: SMOID LLRVD
Benchmarks: SMOID LLRVD
Metrics: PSNR SSIM ST-RRED Params FLOPs
Baselines: SMOID RViDeNet FastDVD EMVD LLRVD FloRNN ShiftNet BNN Bireal IRNet ReActNet BTM
Builds on: ShiftNet Bi-Real Net ReActNet XNOR-Net BBCU
Closest papers by meaning
- Enhancing Low-Light Images: A Synthetic Data Perspective on Practical and Generalizable Solutions (AAAI 2025 · similarity 0.85)
- Contrastive Learning for Low-Light Raw Denoising (Student Abstract) (AAAI 2024 · similarity 0.85)
- Zero-Shot Low-Light Image Enhancement via Latent Diffusion Models (AAAI 2025 · similarity 0.84)
- DPLUT: Unsupervised Low-light Image Enhancement with Lookup Tables and Diffusion Priors (AAAI 2025 · similarity 0.83)
- Binarized Neural Network for Multi-spectral Image Fusion (CVPR 2025 · similarity 0.83)
- Seeing Dark Videos via Self-Learned Bottleneck Neural Representation (AAAI 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.