Research map / Image generation / Image restoration
BurstDeflicker: A Benchmark Dataset for Flicker Removal in Dynamic Scenes
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.
Introduces BurstDeflicker, the first multi-frame flicker removal benchmark, combining Retinex-based synthetic flicker, 4,000 real static burst pairs, and 3,690 green-screen dynamic pairs, and benchmarks Restormer, Burstormer, HDRTransformer, Retinexformer and DeflickerCycleGAN on it.
Problem
Flicker removal for short-exposure images under AC-powered lighting in dynamic scenes is hampered by the lack of large-scale, realistic paired datasets, since aligned flicker/flicker-free pairs with motion are nearly impossible to capture.
Approach
A Retinex-based synthesis that models ambient and flicker illumination and restores flicker to an effective illumination value rather than removing it entirely, combined with real static captures and a green-screen compositing method that adds motion to real flickering burst pairs while preserving real flicker.
Key results
- Restormer trained on BurstDeflicker achieves 30.634 PSNR and 0.045 LPIPS on static test data, outperforming Burstormer (29.439) and HDRTransformer (30.031).
- Adding BurstDeflicker-G green-screen data to synthetic and static data raises dynamic-set MUSIQ from 58.431 to 59.097 and lowers PIQE from 36.259 to 34.896.
- Burst-3 input improves static PSNR by 3.324 dB over single-image input (27.310 to 30.634) with Restormer.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train with synthetic data in addition to BurstDeflicker-S (ablation) | Restormer · BurstDeflicker-S static test | PSNR: 30.481 -> 30.645 | Helped |
| Train with synthetic data in addition to BurstDeflicker-S (ablation) | Restormer · Dynamic test data | MUSIQ: 58.431 -> 59.097 | Helped |
| Add BurstDeflicker-G training data (ablation) | Restormer trained on synthetic data and BurstDeflicker-S · Dynamic test data | MUSIQ: 58.431 -> 59.097 | Helped |
| Add BurstDeflicker-G training data (ablation) | Restormer trained on synthetic data and BurstDeflicker-S · BurstDeflicker-S static test | PSNR: 30.645 -> 30.634 | No clear effect |
| Use 2 input frames instead of a single image (ablation) | Restormer · BurstDeflicker-S static test | PSNR: 27.310 -> 30.264 (+2.594) | Helped |
| Use 3 input frames instead of a single image (ablation) | Restormer · BurstDeflicker-S static test | PSNR: 27.310 -> 30.634 (+3.324) | Helped |
| Increase input from 2 to 3 frames (ablation) | Restormer · BurstDeflicker-S static test | PSNR: 30.264 -> 30.634; 0.730 dB | Helped |
| Use 3 input frames instead of a single image (ablation) | Restormer · BurstDeflicker-S static test | LPIPS: 0.069 -> 0.045 (-0.024) | Helped |
| Use 3 input frames instead of a single image (ablation) | Restormer · Dynamic test data | MUSIQ: 58.664 -> 59.097 | Helped |
| Train on BurstDeflicker instead of the original training setup | Lin et al. model · BurstDeflicker-S static test | PSNR: 20.358 -> 26.408 | Helped |
| Train on BurstDeflicker instead of the original training setup | Retinexformer · BurstDeflicker-S static test | PSNR: 15.704 -> 27.212 | Helped |
| Use Restormer rather than HDRTransformer | MFFR models trained on BurstDeflicker · BurstDeflicker-S static test | PSNR: 30.031 -> 30.634 | Helped |
Surprising findings
- Pretrained Retinexformer (low-light enhancement model) performs poorly on flicker removal, indicating low-light training alone is insufficient.
- Training only on the Retinex-synthetic data gives 24.483 PSNR, well below training on real static data, although the synthetic data helps when combined.
Limitations
- Severe flicker under AC light as sole illumination causes color shifts in restorations.
- Multi-frame methods struggle when available frames lack complete scene content and degrade to single-frame performance under strong handheld jitter.
Experimental setup
Datasets: indoorCVPR VideoMatte240K BurstDeflicker-S BurstDeflicker-G
Benchmarks: BurstDeflicker-S static test set BurstDeflicker dynamic test set (50 handheld sequences)
Metrics: PSNR SSIM LPIPS MUSIQ PIQE BRISQUE
Baselines: Retinexformer DeflickerCycleGAN Burstormer HDRTransformer
Builds on: Retinex theory DeflickerCycleGAN VideoMatte240K Restormer
Closest papers by meaning
- Bricker to BRACE: A Bracket Exposure RAW Dataset and Restoration Model for Flicker-Banding (NeurIPS 2026 · similarity 0.88)
- RIFLE: Removal of Image Flicker-Banding via Latent Diffusion Enhancement (NeurIPS 2026 · similarity 0.86)
- LuSh-NeRF: Lighting up and Sharpening NeRFs for Low-light Scenes (NeurIPS 2024 · similarity 0.82)
- SpikeNeRF: Learning Neural Radiance Fields from Continuous Spike Stream (CVPR 2024 · similarity 0.82)
- Transient Glimpses: Unveiling Occluded Backgrounds through the Spike Camera (AAAI 2024 · similarity 0.82)
- Towards Robust Event-guided Low-Light Image Enhancement: A Large-Scale Real-World Event-Image Dataset and Novel Approach (CVPR 2024 · similarity 0.82)
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.