Research map / Label-efficient and robust learning / Domain adaptation and generalization
Exploring the Common Appearance-Boundary Adaptation for Nighttime Optical Flow
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 ABDA-Flow, a common appearance-boundary adaptation framework that transfers daytime appearance (via reflectance-aligned space) and event boundary knowledge (via spatiotemporal-gradient-aligned space) to nighttime optical flow estimation.
Problem
Nighttime optical flow suffers from weakened texture and amplified noise, which break brightness and gradient constancy and cause invalid motion feature matching; direct domain adaptation from auxiliary domains suffers from feature distribution misalignment.
Approach
Instead of adapting directly in input visual or output motion space, build two intermediate common-latent spaces: a retinex reflectance-aligned space for transferring global motion appearance from daytime images, and a spatiotemporal-gradient-aligned space for contrastively transferring local boundary knowledge from accumulated events.
Key results
- On synthetic D-KITTI, ABDA reaches EPE 3.47 and Fl-all 16.13%, versus 8.42 EPE for the best competing unsupervised variant (SMURF + AGLLNet+).
- On real Dark-DSEC, ABDA achieves EPE 0.74 and Fl-all 11.85%, outperforming DarkFlow-PWC (2.29 EPE) and GyroFlow.
- Ablation on common space shows EPE dropping from 1.51 (no common space) to 0.74 with motion, reflectance and boundary adaptation combined.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use common appearance-boundary adaptation instead of DarkFlow-PWC | nighttime optical flow · D-KITTI | EPE: 7.56 -> 3.47 | Helped |
| Use common appearance-boundary adaptation instead of DarkFlow-PWC | nighttime optical flow · ND-KITTI | Fl-all: 41.28% -> 23.86% | Helped |
| Use ABDA instead of GyroFlow | nighttime optical flow · Dark-GOF | EPE: 0.92 -> 0.85 | Helped |
| Use ABDA instead of E-RAFT | nighttime optical flow · Dark-DSEC | EPE: 0.82 -> 0.74 | Helped |
| Add motion adaptation to the setting without motion, reflectance, or boundary adaptation (ablation) | ABDA · Dark-DSEC | EPE: 1.51 -> 1.09 | Helped |
| Add reflectance adaptation (ablation) | ABDA with motion adaptation · Dark-DSEC | EPE: 1.09 -> 0.87 | Helped |
| Add boundary adaptation (ablation) | ABDA with motion adaptation · Dark-DSEC | EPE: 1.09 -> 0.95 | Helped |
| Add boundary adaptation to motion and reflectance adaptation (ablation) | ABDA · Dark-DSEC | EPE: 0.87 -> 0.74 | Helped |
| Add intra-domain alignment loss (ablation) | ABDA adaptation losses · Dark-DSEC | EPE: 1.45 -> 1.24 | Helped |
| Add inter-domain alignment loss (ablation) | ABDA adaptation losses · Dark-DSEC | EPE: 1.24 -> 1.05 | Helped |
| Add contrastive loss (ablation) | ABDA adaptation losses · Dark-DSEC | EPE: 1.05 -> 0.85 | Helped |
| Add self-flow consistency loss (ablation) | ABDA adaptation losses · Dark-DSEC | EPE: 0.85 -> 0.74 | Helped |
Surprising findings
- Under the ablation, CycleGAN-based multi-source training with E-RAFT performs worse than common space adaptation alone, suggesting training data diversity alone is insufficient.
Limitations
- Requires paired nighttime image and event data, obtained via a coaxial optical system or stereo rectification with registration fine-tuning.
- Relies on manually defined motion classes K determined by correlation statistics, a hand-designed component.
Experimental setup
Datasets: KITTI2015 Dark-KITTI2015 Dark-GOF Dark-DSEC LLFE
Benchmarks: Dark-KITTI2015 Dark-GOF Dark-DSEC LLFE
Metrics: EPE Fl-all
Baselines: DarkFlow-PWC GyroFlow PWC-Net GMA Selflow SMURF KinD++ AGLLNet EV-FlowNet E-RAFT CycleGAN
Builds on: Uretinex-Net RAFT SMURF GyroFlow DarkFlow
Closest papers by meaning
- NightCC: Nighttime Color Constancy via Adaptive Channel Masking (CVPR 2024 · similarity 0.87)
- Rethinking Nighttime Image Deraining via Learnable Color Space Transformation (NeurIPS 2025 · similarity 0.87)
- See through the Dark: Learning Illumination-affined Representations for Nighttime Occupancy Prediction (NeurIPS 2025 · similarity 0.86)
- NightRain: Nighttime Video Deraining via Adaptive-Rain-Removal and Adaptive-Correction (AAAI 2024 · similarity 0.85)
- CEDFlow: Latent Contour Enhancement for Dark Optical Flow Estimation (AAAI 2024 · similarity 0.85)
- NightAdapter: Learning a Frequency Adapter for Generalizable Night-time Scene Segmentation (CVPR 2025 · similarity 0.85)
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.