The aim of the thesis is to propose and evaluate an algorithm for image segmentation, focussing on performance on a large set of classes without pixel-accurate training masks. However, for the unseen classes, a convolutional deep net classifier is available.

For supervised image segmentation, annotated images are required. Labels for segmentation are typically obtained by manually annotating each pixel. Annotating a large dataset for segmentation can be a time-consuming and expensive task. On the other hand, classification convolutional neural networks are usually trained using class labels. In general, class labels for image classification are easier and cheaper to obtain than pixel-level annotated images for segmentation.

The student will use a classification neural network as an encoder and a class-agnostic model as a decoder for image segmentation. From the classification convolutional neural network (encoder), we can easily obtain so-called class activation maps. These activation maps could be understood as an approximation of segmentation, and the decoder can get them as input.

In this thesis, the student will:
- Collect datasets necessary for the thesis.
- Design a class-agnostic segmentation method.
- Train a segmentation network that takes class activation maps, obtained from an available classification network, as input.
- Train a combined method, using both an image and class activation maps as input.
- Compare segmentation results on known and unknown classes.
- Propose a method that aims at outperforming the combined segmentation method, on unknown classes.
- Write up the results in a thesis form.