BurstDeflicker: A Benchmark Dataset for Flicker Removal in Dynamic Scenes

TL;DR

BurstDeflicker tackles flicker artifacts caused by rolling shutter exposure under AC lighting by introducing the first multi-frame flicker removal (MFFR) benchmark. It combines three data sources: a Retinex-based synthetic flicker generator with $I_{flicker} = R \odot (L_a + L_f)$ and $I_{clean} = R \odot (L_a + \overline{L_f})$, 4,000 real static flicker images, and 3,690 green-screen dynamic sequences to cover synthesis, realism, and motion. The authors provide training pipelines, pretraining on synthetic data, and a green-screen motion-embedding approach to mitigate motion ghosts, with extensive experiments showing improved performance, particularly when using 3-frame inputs. The dataset is positioned to advance flicker removal research and improve robustness in real-world, dynamic environments, benefiting high-speed photography, HDR imaging, and video capture under artificial lighting.

Abstract

Flicker artifacts in short-exposure images are caused by the interplay between the row-wise exposure mechanism of rolling shutter cameras and the temporal intensity variations of alternating current (AC)-powered lighting. These artifacts typically appear as uneven brightness distribution across the image, forming noticeable dark bands. Beyond compromising image quality, this structured noise also affects high-level tasks, such as object detection and tracking, where reliable lighting is crucial. Despite the prevalence of flicker, the lack of a large-scale, realistic dataset has been a significant barrier to advancing research in flicker removal. To address this issue, we present BurstDeflicker, a scalable benchmark constructed using three complementary data acquisition strategies. First, we develop a Retinex-based synthesis pipeline that redefines the goal of flicker removal and enables controllable manipulation of key flicker-related attributes (e.g., intensity, area, and frequency), thereby facilitating the generation of diverse flicker patterns. Second, we capture 4,000 real-world flicker images from different scenes, which help the model better understand the spatial and temporal characteristics of real flicker artifacts and generalize more effectively to wild scenarios. Finally, due to the non-repeatable nature of dynamic scenes, we propose a green-screen method to incorporate motion into image pairs while preserving real flicker degradation. Comprehensive experiments demonstrate the effectiveness of our dataset and its potential to advance research in flicker removal.

Figures (8)

• Figure 1: (a) Illustration of flicker formation. The moving object is illuminated by stable light and AC-powered flickering sources. The intensity of the flickering component changes over time (purple area), and each row is exposed at a slightly different moment, leading to a non-uniform brightness distribution across the captured image. (b) Capturing short-exposure images under artificial lighting often results in flicker degradation (red box in the left). Although increasing the exposure time can mitigate flicker artifacts, it introduces motion blur pan1pan2chen2025polarization (red box in the middle). Our method effectively removes flicker while preserving fine image details (green box in the right).
• Figure 2: A visual illustration of the three-stage growth of the BurstDeflicker dataset.
• Figure 3: Flicker synthesis results based on the proposed Retinex-based method. Background images are sourced from the indoorCVPR dataset cvprindoor. A pre-training is conducted using the synthetic data, providing a strong initialization for subsequent training on real data.
• Figure 4: Illustration of the real captured dataset. (a) Example images from our dataset, which include a diverse range of common artificial lighting scenarios. (b) The intensity distributions of flicker and non-flicker frames. (c) The area ratio of flicker degradation per image. (c) The luminance distribution of flickering and clean images across different scenes.
• Figure 5: The motion synthesis process and training pipeline. (a) The green-screen footage is selected from the VideoMatte240K dataset green-screen. The compositing of green-screen foregrounds with the backgrounds is manually performed using Adobe After Effects. (b) Given a sequence of flickering frames, we select three frames as input to form a training batch, with the target being a single clean reference image. The synthetic dataset and BurstDeflicker-S also follow this pipeline.