As autonomous driving is quickly becoming a reality, the most considerable challenge with self-driving remains safety, as the car must process surroundings to make the correct predictions. This thesis delves into the detection and classification of dynamic objects in traffic, such as vehicles and pedestrians, on a pixel level. Two experiments are presented, distinguished only by the data used during the training process: one utilizing a front-facing camera and the other a combination of front and left-facing cameras to capture a diagonal view. Data collection relied on the CARLA simulator, purposely omitting standard technologies like lidar or radar, and focusing exclusively on monocular data. The objective of this research is to examine the generalization capabilities of models trained on restricted perspectives when exposed to inputs from cameras capturing a full 360° view of their surroundings. The described challenge is tackled with the task of instance segmentation using the most famous instance segmentation algorithms: Mask Region-based Convolutional Neural Network and YOLOv7. The results show to what extent a restricted monocular setup can be used for the instance segmentation of the perspectives that differ from the perspectives included in the training phase.