BlurPoint: Efficient Motion Blur Aware Student-Teacher Local Feature Learning
Wenting Wang, Zhenjun Zhao, Jiaxin Guo, Yunhui Liu, Charlie C.L. Wang, Yeung Yam
AI summary
Problem
Most local feature detectors degrade under motion blur, and existing two-stage deblur-then-extract approaches are computationally expensive and prone to artifacts. There is a lack of efficient, one-stage methods that can learn robust features directly from blurred images without ground-truth labels.
Approach
A student-teacher framework where a frozen teacher model pre-trained on sharp images guides a student model trained on blurred images via feature divergence and triplet knowledge distillation losses, all within a self-supervised learning setup.
Key results
- State-of-the-art accuracy on homography and relative pose estimation for blurred images
- Competitive performance on sharp images maintained
- Superior efficiency-performance balance over deblur-then-match pipelines
- Effective knowledge transfer from sharp to blurred domains without labeled data
Why it matters
Enables reliable 3D vision and robotics tasks like SLAM and visual localization in challenging low-light or high-motion environments where motion blur is common.
Abstract
Local feature detection and description serve as the foundation for many 3D vision tasks. However, most existing algorithms rely on sharp images, resulting in degraded performance when motion blur occurs due to long exposure. To tackle this challenge, we propose an effective end-to-end model that jointly learns feature detection and description from blurred images in a self-supervised manner, without requiring any additional labeled data. Rather than simply mixing sharp and blurred samples during training, we design a student–teacher framework to explicitly transfer knowledge from sharp to blurred domains. The teacher model extracts local features from sharp images and enforces photometric consistency in feature space, which is then distilled to the student model trained on blurred inputs. To facilitate this knowledge transfer, we introduce two tailored loss functions, feature divergence loss and triplet knowledge distillation loss, both aimed at aligning feature representations under motion blur. Extensive experiments on homography estimation, relative pose estimation, and visual localization demonstrate that our method achieves state-of-the-art performance on blurred im- ages, while maintaining competitive accuracy on sharp images.