Challenges: - How do you run a neural net over 100 million video frames without spending $1M? - How do you manage 20 TB of video data for both analytics and visualization? - How do you write queries over space, time, and language? - How do you even Kubernetes?
-
-
Show this thread
-
- How do you close the loop of: run ML model, visualize outputs, label fixes, re-train, repeat? - How can you propagate type information throughout the system? e.g. if I have a bounding box or 2D pose keypoints, then that should imply how to store it, visualize it, analyze it, ..
Show this thread -
Lessons learned: - Jupyter widgets are underrated. You can easily build stateful workflows that go between Python data science code <-> Javascript frontend code (e.g. active learning loop). See https://willcrichton.net/notes/rapid-prototyping-data-science-jupyter/ … - Use docker for everything, even local development.
Show this thread -
- Every time you 10x the data, a new and unexpected bottleneck will appear. It never just scales like you want. - Invest in learning a Python <-> C++/Rust interface. Preferably Rust, because I deadlocked the GIL with C++ since it **implicitly wraps static init in a mutex!!!!!**
Show this thread -
And on a more somber note: work on prosocial applications. There's a reason we worked on transparency in news media, and NOT surveillance. Sadly, that makes us an outlier in the systems-for-video community.
Show this thread
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.
cognitive psychology. PhD