horsellama

Regarding AI, the thing I’ve heard the most lately is “during a gold rush sell shovels”.

It’s a great advice (?) for a semiconductor company ready to get into the gpu market and compete with Nvidia, but what about me? I’m just a machine learning engineer bored to death by data cleaning and writing over and over the same two lines of PyTorch. I guess in that metaphor I’m the one using such shovels.

Perhaps building shovels is much more interesting than using them? What if I’d want to jump into CUDA kernel development? Is that a niche hard enough that the competition is low? Say I’m willing to put in the time and effort, where should I start? Best resources on the topic?

FYI, I’m well versed in C++ but zero in GPU programming

Say I have a video from a random camera somewhere in the world. The camera view is fixed and we don’t know neither its intrinsic nor extrinsic parameters.

Is there a way of automatically place 3D objects in the scene? That is automatically find a ground plane and somehow calibrate it using reference objects.

Is this scalable? Say I have thousands of different videos, now any manual step is not feasible. Is there a way of using the video itself for this calibration problem?

Any recommendation for a small PC CUDA capable? I don’t want to train or run inference of large models, just do some cuda and PyTorch prototyping.