Read on Twitter
Wide baseline stereo PhD blog: Lessons learned and future directions.
https://ducha-aiki.github.io/wide-baseline-stereo-blog/2020/11/26/lessons-and-future-directions.html
Tl;dr in the thread below.
Lessons.
1. Benchmark = metrics + implementation + dataset. Dataset ultimately defines what you can or cannot measure
1/7
https://ducha-aiki.github.io/wide-baseline-stereo-blog/2020/11/26/lessons-and-future-directions.html
Tl;dr in the thread below.
Lessons.
1. Benchmark = metrics + implementation + dataset. Dataset ultimately defines what you can or cannot measure
1/7
2. Trying to work for the "most general" case might be detrimental for the practical applications. Example: upright features vs orientation detection
3. When borrow idea from classical paper, adapt it
4. Classic handcrafted algorithms are not dead and can be improved a lot.
2/7
3. When borrow idea from classical paper, adapt it
4. Classic handcrafted algorithms are not dead and can be improved a lot.
2/7
Future research directions
1. Application-specific local features.
First steps: SIPs( @tcies, @CSProfKGD, @davsca1 ), HOWs ( @tojenicek, @giotolias) , Reinforced SuperPoint ( @eric_brachmann)
3/7
1. Application-specific local features.
First steps: SIPs( @tcies, @CSProfKGD, @davsca1 ), HOWs ( @tojenicek, @giotolias) , Reinforced SuperPoint ( @eric_brachmann)
3/7
2. Rethinking the WBS pipeline for a specific application.
Examples: works by @AjdDavison, for SLAM, @pesarlin (Hierarchical Localization, SuperGlue), our new paper which I will promote on Monday
4/7
Examples: works by @AjdDavison, for SLAM, @pesarlin (Hierarchical Localization, SuperGlue), our new paper which I will promote on Monday
4/7
3. Rethinking and improving the process of training data generation for the WBS.
So far we have (a) SfM data (b) affine and color augmentation (c) synthetic data.
(b) limits are reached, (a) is not always applicable. We need more of (c) or maybe something new
5/7
So far we have (a) SfM data (b) affine and color augmentation (c) synthetic data.
(b) limits are reached, (a) is not always applicable. We need more of (c) or maybe something new
5/7
4. Matching with On-Demand View Synthesis revisited.
We have affine view synthesis (MODS, ASIFT), GAN-based day-night translations and that's all. We need more
Examples: "MonoDepth Helps Matching" ( @dantkz ,, @SattlerTorsten), "Single view Calibration" ( @ylochman , @prittjam ) 6/7
We have affine view synthesis (MODS, ASIFT), GAN-based day-night translations and that's all. We need more
Examples: "MonoDepth Helps Matching" ( @dantkz ,, @SattlerTorsten), "Single view Calibration" ( @ylochman , @prittjam ) 6/7
5. Moar (hallucinated) inputs to the WBS! I have already mentioned monodepth. Semantic segmentation? Yes, please. Surface normals? Why not? Intrinsic images? Йой, най буде!
That's it, folks!
7/7
That's it, folks!
7/7
