Dynamic Exposure Burst Image Restoration
Burst image restoration in low-light conditions typically relies on fixed exposure settings that limit complementary information across frames. This paper proposes DEBIR, a pipeline that dynamically predicts per-frame exposure times using a Burst Auto-Exposure Network (BAENet) conditioned on preview images, motion, and gain. The key insight is that scene-adaptive exposures can optimally trade off noise and blur across the burst, and the authors enable end-to-end training via a novel differentiable burst simulator that eliminates the need for ground-truth exposure sequences.
The paper addresses a compelling gap in burst photography—optimal exposure control—with solid technical contributions including a differentiable burst simulator and a careful three-stage training strategy. The method demonstrates consistent improvements over fixed bracketing and classical auto-exposure. However, the evaluation relies heavily on synthetic training data generated via 8× frame interpolation from 240 FPS video, and real-world validation is limited to no-reference metrics without ground truth, making it difficult to assess true restoration accuracy.
The differentiable burst simulator is technically sound and enables the core training mechanism without requiring ground-truth exposure sequences. The three-stage training strategy (pre-training restoration, warm-up with pseudo-ground-truth exposures, then end-to-end fine-tuning) effectively handles the instability of joint optimization. Ablation studies thoroughly validate design choices: non-uniform exposures improve PSNR by 0.31 dB over uniform exposures, and each input component (preview image, gain, motion) contributes meaningfully to performance.
The training data synthesis assumes high-FPS video (1920 FPS) obtained via frame interpolation, which may not accurately model real sensor noise and motion blur characteristics. The method depends on optical flow estimates between preview frames for motion magnitude, creating a potential failure mode for fast motion. While Table 1 reports gains over exposure bracketing, the margin is modest (0.28 dB PSNR), and the real-world evaluation in Table 2 relies solely on no-reference metrics (NIQE, BRISQUE, TOPIQ) for 142 captured bursts, which are controversial proxies for actual restoration quality.
The controlled comparisons are fair: all methods use the same restoration backbone (Burstormer) and training protocol. DEBIR reasonably outperforms Digital-Gimbal (exposures fixed after training) and Active S-L (limited to two images). However, the main results show diminishing returns compared to fixed exposure bracketing (35.32 dB vs 35.04 dB), raising questions about the practical impact of the added complexity. The paper does not compare against oracle or exhaustive search baselines to quantify the suboptimality gap of BAENet's predictions.
The paper provides detailed implementation information including network architectures (MobileNetV2, Burstormer), hyperparameters ($t_u=128/1920$ sec, learning rates $3\times10^{-4}$, $1\times10^{-7}$, $1\times10^{-5}$), and dataset construction from GoPro and RealBlur. The bounded softmax formulation for exposure prediction is clearly specified. However, no code or pre-trained models are released, and the supplementary material containing additional dataset details and camera specifications is not available in the provided text, limiting immediate reproducibility.
Burst image restoration aims to reconstruct a high-quality image from burst images, which are typically captured using manually designed exposure settings. Although these exposure settings significantly influence the final restoration performance, the problem of finding optimal exposure settings has been overlooked. In this paper, we present Dynamic Exposure Burst Image Restoration (DEBIR), a novel burst image restoration pipeline that enhances restoration quality by dynamically predicting exposure times tailored to the shooting environment. In our pipeline, Burst Auto-Exposure Network (BAENet) estimates the optimal exposure time for each burst image based on a preview image, as well as motion magnitude and gain. Subsequently, a burst image restoration network reconstructs a high-quality image from burst images captured using these optimal exposure times. For training, we introduce a differentiable burst simulator and a three-stage training strategy. Our experiments demonstrate that our pipeline achieves state-of-the-art restoration quality. Furthermore, we validate the effectiveness of our approach on a real-world camera system, demonstrating its practicality.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.