Juho Kannala
Papers
11
Total Citations
156
H-Index
5
About
Juho Kannala is a leading researcher in computer vision and robotics, with a primary focus on visual localization, camera relocalization, and 3D scene understanding. His most significant contributions center on developing deep learning methods for estimating camera pose from single images, a critical capability for augmented reality, autonomous navigation, and robotics. Kannala pioneered the concept of full-frame scene coordinate regression, where convolutional neural networks directly predict 3D world coordinates for every pixel, enabling robust pose estimation without explicit 3D models. His 2018 paper on this topic has garnered 38 citations, while his subsequent work on hierarchical scene coordinate classification and regression (HSCNet++), incorporating transformer architectures, represents the state of the art in visual localization. Kannala has also advanced the field through angle-based reprojection losses that improve geometric consistency, and his work on few-shot scene region classification addresses practical deployment challenges. Beyond localization, he contributed to multi-sensor reconstruction with the MuSHRoom dataset and developed RealAnt, an open-source, low-cost quadruped robot platform for reinforcement learning research. His work consistently bridges theoretical innovation with practical robotics applications, earning recognition across computer vision and robotics communities.
Research Focus
Key Achievements
Top Papers
- 1Full-Frame Scene Coordinate Regression for Image-Based Localization38 citations · 2018
- 2Visual Localization via Few-Shot Scene Region Classification33 citations · 2022
- 3
- 4
- 5
- 6
- 7Expansion of Visual Hints for Improved Generalization in Stereo Matching4 citations · 2023
- 8Full-Frame Scene Coordinate Regression for Image-Based Localization3 citations · 2018
- 9
- 10