Detection draws a box around an object. Segmentation goes down to the pixel: it decides which pixels belong to which object or region. That precision matters for medical scans, self-driving, photo editing, and robotics.
There are three main flavours. Semantic segmentation labels every pixel with a class, instance segmentation separates individual objects of the same class, and panoptic segmentation combines both. U-Net is the classic encoder-decoder network for dense prediction, and Mask R-CNN extends a detector with a mask branch.
Recently, promptable foundation models such as the Segment Anything Model can segment objects they were never trained on, given a point or a box. Watch the lessons in order, then continue on to Vision-Language & Generative Models.
How the main segmentation techniques differ, with DeepLab v3+, Mask R-CNN, and UPSNet as reference models.
The U-Net architecture step by step: the encoder path, the bottleneck, the decoder path, and how skip connections carry features across.
An introduction to Mask R-CNN: how instance segmentation differs from detection, the architecture, the region proposal network, and the limitations of the model.
The promptable foundation model from Meta AI: why it was built, how it works, the SA-1B dataset, and zero-shot results.