Most modern machine learning techniques based on neural networks need large amounts of task-specific and annotated data to achieve good performance. However, it is tedious to obtain instance-level annotations in real-world settings and is a challenge in many applications of neural networks. This master thesis explores the task of learning with limited
labels for object detection. More specifically, we would like to address the task of classifying and regressing the position of the most prominent foreground object in an image.

The proposed method uses a two-stage pipeline. In the first stage, the goal is to localize the salient object in an image in a class-agnostic manner. To accomplish this, we leverage the advances in self-supervised learning techniques that can provide descriptive features. These feature maps are effectively processed to obtain precise bounding boxes over the salient object.

In the second stage of the pipeline, the task is to classify the discovered salient objects into different semantic categories. The proposed method uses techniques based on multi-modal text-visual and self-supervision learning schemes with limited labels.