“Yeah It’s on. But wait, it doesn’t look good.”
Motion Blur
Motion blur could cause lower number of features and feature mismatching. These will further cause huge errors in PnP solving. Below is a scene of my room. Both images are blurred. See how many feature matches are missed and mismatched?
Debluring
In Roh et al.’s MBA-VO paper (ICCV 2021), two deep learning-based deblurring methods were evaluated:
1
2
- A slower RNN with better quality [Tao et al., 2018](https://openaccess.thecvf.com/content_cvpr_2018/papers/Tao_Scale-Recurrent_Network_for_CVPR_2018_paper.pdf).
- A faster, but slightly less accurate, CNN (Kupyn et al., 2019).
For my RGBD SLAM, I chose the multiscale RNN due to its effectiveness in offline applications. This method creates an image pyramid, applies deblurring at multiple resolutions, and avoids overfitting through weight sharing. - Tao et al. found that CNN with independent params could overfit.
