Image segmentation, a crucial aspect of computer vision, involves dividing an image into segments where each pixel is associated with an object type. The two primary types are semantic segmentation, marking all objects of the same type with one class label, and instance segmentation, giving similar objects separate labels. The basic architecture for image segmentation is an encoder-decoder model, with notable architectures like U-Net, FastFCN, Gated-SCNN, DeepLab, and Mask R-CNN each offering unique methods for handling segmentation tasks. Loss functions such as focal loss, dice loss, and boundary loss are used to enhance accuracy by addressing class imbalance or focusing on hard examples. Datasets like COCO, PASCAL VOC, and Cityscapes provide the necessary data for training these models, while frameworks like FastAI, OpenCV, and MIScnn facilitate their implementation. The article also highlights the use of Neptune for tracking and comparing model performance, emphasizing the importance of experiment tracking in the development of image segmentation models.