Semantic segmentation is the machine learning task of learning a class label for each pixel of image data. This approach has been useful across many domains, from Self-Driving Cars to Medical Diagnostics.
Despite the growing popularity of this technique, the tools for debugging these models have lagged behind. To help bridge that gap, Weights & Biases is releasing native support for semantic segmentation. Instead of recording your results as static images, you can now log your segmentation maps directly via our logging API. You can analyze your results interactively in our UI, which lets you toggle mask types (e.g. prediction versus ground truth) and classes (car, road, etc) on or off and change the opacity of these items. This makes it easier to explore the details of your experiments and save the results of your exploration for sharing with collaborators, helping them understand your work faster through a visualization specific to the problem you are tackling.