Data Preprocessing and the Albumentations Library

Segmentation datasets, like object detection datasets, require a large upfront time investment. With segmentation datasets though, you'll typically be annotating everything in an image, instead of just objects of interest, and you'll be doing so more accurately along the borders of the object, instead of a box around it.

Because of this, you're likely going to be working with in-house segmentation datasets, labelled by a team trying to solve a particular problem. Segmentation models are more commonly applied to specific use-cases, and trying to train a general segmentation model, without a niche use in mind is rarer. Semantic segmentation models are more sensitive to domain shift than image classification models, in large part because image classification models are blind to many small differences by abstracting them away, while segmentation models pay much more attention to small details. The good thing is - since you're likely going to be making it for a niche problem, you'll have the same type of input during and after training!

Just like with classification and detection - datasets and models are typically single-class or multi-class. Single-class would be detecting "traffic sign" in an image or segmenting it. Multi-class would be detecting "stop sign", "crosswalk" and "wrong way" signs, or segmenting them.