2024

2024-05-01

Agenda

• Discussion about kornia 2024 roadmap.

Minutes

• We ended the year with >2M downloads per month: https://www.pepy.tech/projects/kornia 🥳
• GSOC 2024: let's keep trying! application period opens Jan 22
  • https://summerofcode.withgoogle.com/
• General Roadmap 2024: https://docs.google.com/document/d/1zJGDb5Tguri6phR5XVLIgyfR9OGqI-8nQ1kgx4BFEBI/edit?usp=sharing
  • list what the core team commit to take ownership
  • we will assist to land any other proposed feature that makes sense for project

Infrastructure:

• Release schedule: one release per quarter (4 releases per year).
  • By december 2024 release v0.8
  • Q1:January, February, and March (ends on March 31st)
  • Q2: April, May, and June (ends on June 30th)
  • Q3: July, August, and September (ends on September 30th)
  • Q4: October, November, and December (ends on December 31st)
• Rebase master -> main:
  • Do it for the first release of the year and co-exist with the master.
  • Cons: conflict with existing open PRs. Close/Stale open PRs that are older than a couple months (?) approx.
• (Sponsorship) Ask for community support in case older versions of torch are required.
  • Add into the BACKERS/SPONSORS docs
  • Mention also in the development guidelines.
  • Google Summer of Code 2024 (need a separated meeting for that).
  • Update github sponsors profile: https://github.com/sponsors/kornia
• Make documentation more dynamic:
  • Add github discussions
  • Add more dynamic plots
• Google points to old page of tutorials -> check google analytics (?)

New features

• kornia.optim: minimum non linear solvers needed for vision and robotics tasks.
• kornia.tracking: finish pyramidal optical flow: https://github.com/kornia/kornia/pull/2658
• kornia.sensors:
  • Match CameraModel with Sophus
  • Add Imu with pre integration
  • Add Gps support data structure
  • Consider Lidar for point cloud manipulation
• Kornia data structures:
  • refactor all the geometry data structures
  • https://github.com/kornia/kornia/issues/2691
• Face Detection:
  • Update the existing face detection algorithm:
  • https://github.com/kornia/kornia/issues/1781
  • Make sure we can export to the edge via onnx/torchscript/compile
• People detection:
  • identify a small model similar to yunet that can work fast on cpu
• Monocular depth estimation:
  • Find a state of the art monocular depth estimation network and support it.
  • Add utilities for depth estimation.
  • Potential methods:
  • https://github.com/aim-uofa/AdelaiDepth (not good license for kornia integration)
  • https://github.com/isl-org/ZoeDepth (not good enough, yet)
  • https://github.com/YvanYin/Metric3D
• Deep Dense Stereo Optical flow:
  • Find a state of the art network to compute optical flow and support it.
  • Add utilities for stereo / or optical flow

Integrations

• Keep an eye into Keras 3.0:
  • they have good models like SAM'
  • As GSOC ?

Action items

• setup releases milestones in github @joao
• send calendar invites for releases to the core team @edgar
• write formal roadmap and convert to issues @edgar
• organize a meeting to discuss GSOC before Jan 22nd

<><><><><><><><><><><><><><><><><><><><><><><><><><><><><><><>