Research map / 3D vision / Depth and stereo
Dual Exposure Stereo for Extended Dynamic Range 3D Imaging
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 dual-exposure stereo that alternates two exposures under an auto dual exposure control (ADEC) and fuses motion-compensated dual-exposure features in RAFT-Stereo to extend the effective dynamic range of stereo depth estimation.
Problem
Stereo depth estimation fails under extreme illumination because the limited dynamic range of cameras leaves regions under- or over-exposed, and existing auto-exposure or fixed-bracketing approaches do not expand the native camera DR for 3D imaging.
Approach
Synchronized stereo capture alternates two exposures that adaptively diverge or converge based on histogram statistics (skewness and extreme-pixel ratios), and a motion-aware, weighted feature-level fusion of the two frames' stereo features is used to estimate disparity over a wider effective DR.
Key results
- ADEC reduces synthetic disparity MAE to 1.3837 px versus 1.7198 (NeuralAE), 2.2958 (AverageAE), and 2.4680 (GradientAE).
- On real data, ADEC achieves depth MAE of 1.6518 m versus 1.7565 (NeuralAE), 2.6677 (GradientAE), and 2.7089 (AverageAE).
- ADEC runs at 124.58 FPS for the exposure control module, versus 42.10 for GradientAE and 0.25 for NeuralAE; the method retains high accuracy up to 160% DR expansion.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ADEC with dual-exposure stereo instead of single-exposure AEC | RAFT-Stereo · Synthetic dataset | disparity MAE [px]: 1.7198 -> 1.3837 | Helped |
| Use ADEC with dual-exposure stereo instead of single-exposure AEC | RAFT-Stereo · Real dataset | depth MAE [m]: 1.7565 -> 1.6518 | Helped |
| Replace RAFT-Stereo [24] with backbone [51] (ablation) | ADEC · Ablation study | disparity MAE [px]: 1.76 -> 1.46 | Helped |
| Use ADEC instead of AverageAE, GradientAE, or NeuralAE (ablation) | RAFT-Stereo [24] · Ablation study | disparity MAE [px]: 2.58, 2.52, 1.92 -> 1.76 | Helped |
| Use ADEC instead of AverageAE, GradientAE, or NeuralAE (ablation) | backbone [51] · Ablation study | disparity MAE [px]: 2.18, 1.79, 1.67 -> 1.46 | Helped |
| Use feature-level fusion instead of exposure-level or disparity-level fusion (ablation) | Dual-exposure stereo · Ablation study | disparity MAE [px]: 3.22 -> 1.82; 5.03 -> 1.82 | Helped |
| Add motion compensation to exposure-level fusion (ablation) | Dual-exposure stereo · Ablation study | disparity MAE [px]: 3.22 -> 2.65 | Helped |
| Add motion compensation to disparity-level fusion (ablation) | Dual-exposure stereo · Ablation study | disparity MAE [px]: 5.03 -> 3.75 | Helped |
| Add motion compensation to feature-level fusion (ablation) | Dual-exposure stereo · Ablation study | disparity MAE [px]: 4.02 -> 1.82 | Helped |
| Add ADEC (ablation) | Dual-exposure stereo · Ablation study | disparity MAE [px]: 6.28 -> 2.90 | Helped |
| Add weighted feature fusion (ablation) | Dual-exposure stereo · Ablation study | disparity MAE [px]: 3.39 -> 2.90 | Helped |
| Add motion compensation (ablation) | Dual-exposure stereo · Ablation study | disparity MAE [px]: 8.37 -> 2.90 | Helped |
Surprising findings
- Neural AEC (NeuralAE) runs at only 0.25 FPS, making it impractical for real-time use despite being the strongest baseline among prior methods.
Limitations
- Dual-exposure fusion assumes overlapping content between alternating frames, so extremely diverged exposures are disallowed.
- Relies on motion between alternating frames being estimated well by optical flow; static or fast-moving scenes may be affected.
- Real-world dataset ground truth is sparse LiDAR-projected depth rather than dense ground truth.
Experimental setup
Datasets: CARLA synthetic stereo dataset Real-world stereo video dataset captured with robot (LiDAR pseudo ground truth)
Benchmarks: Synthetic dual-exposure stereo dataset Real-world stereo video dataset
Metrics: Disparity MAE Depth MAE FPS DR expansion rate
Baselines: AverageAE GradientAE NeuralAE RAFT-Stereo
Builds on: RAFT-Stereo RAFT optical flow CARLA
Closest papers by meaning
- Zero-Shot Event-Intensity Asymmetric Stereo via Visual Prompting from Image Domain (NeurIPS 2024 · similarity 0.84)
- Mono2Stereo: A Benchmark and Empirical Study for Stereo Conversion (CVPR 2025 · similarity 0.83)
- UltraFusion: Ultra High Dynamic Imaging using Exposure Fusion (CVPR 2025 · similarity 0.83)
- All-directional Disparity Estimation for Real-world QPD Images (CVPR 2025 · similarity 0.83)
- Stereo Anywhere: Robust Zero-Shot Deep Stereo Matching Even Where Either Stereo or Mono Fail (CVPR 2025 · similarity 0.83)
- PPMStereo: Pick-and-Play Memory Construction for Consistent Dynamic Stereo Matching (NeurIPS 2025 · 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.