Abstract Affordances play a crucial role in robotics since they allow developing truly autonomous robots, which can freely explore and interact with the environment. Most of the existing approaches for analyzing affordances in a scene consider only one or few types of affordance, e.g., grasping points, object manipulation or locomotion. In many cases only whole objects are considered. In our study we include in total 12 affordances of object-related, manipulation and locomotion affordances, considering affordances of both objects and/or their parts. We design a system that can densely predict affordances given only a single 2D RGB image. For this, we propose a method that transfers object class labels to affordances. This enables us to train convolutional neural networks, a PSPNet-based network and a U-Net-style network, to directly predict affordances from an image using a selective binary cross entropy loss function. The method is able to handle (potentially multiple) affordances of objects and their parts in a pixel-wise manner even in the case of incomplete data. We perform qualitative as well as quantitative evaluations with simulated and real data including robot experiments. In general, we find that frequent affordances are recognized with a substantial fraction of correctly assigned pixels, while this is harder for infrequent affordances and small objects. In addition, we demonstrate that our method performs better than a recent competitive approach. As the proposed method operates on 2D images, it is easier to implement than competing 3D methods and it could therefore more easily provide useful affordance estimates for robotic actions as demonstrated experimentally.