Releasing a new version of a web application or larger software project usually involves some sort of build-deploy-test-fix cycle until it passes QA. A common approach is to have release candidates, however this implies re-building the last RC with the final tag which can be avoided with build numbers.
Recently when converting a PyTorch checkpoint for CPU inference with ONNX throughput fell off a cliff. Of course, runtime on CPU is generally expected to be slower than on GPU, however something was off. Turns out a particular case of floating values was to blame.
Detecting the longest line in an Hough accumulator is easy, however detecting multiple lines is a more annoying problem.
The machine readable zone (MRZ) of identity documents has various checksums which are intended to verify that a readout by a machine is actually correct. But how likely is it that this mechanism fails?
A video is made up of frames, but retrieving a specific frame from a video file is not straight forward. What if the video file is stored remotely?
Docker is awesome. It is even more awesome if you embrace it. So why stop at using it for production images?
This is writeup of some formulas that proved to be helpful understanding and working with homographies.
Swig is really great to wrap native projects. I have been working on wrapping native code in an Android app, figuring out how to run swig as part of the gradle build pipeline took me way too many google queries.
Building a rendering engine comes with a lot of hurdles. One of the problems that might pop out of nowhere (depending on your preparation) is handling meshes.
You know that feeling when you have been using some rule blindly forever and one day you finally spend the time to figure out where this rule comes from? Below is my effort to visualize the integrals which lead to the formulas for surface area and volume of spheres.