PROJECT TITLE :

Deep Hough Transform for Semantic Line Detection

ABSTRACT:

We concentrate on a fundamental task known as semantic line detection in natural scenes, which involves identifying meaningful line structures. A great number of previously proposed solutions consider this issue to be a particular instance of object detection and modify already existing object detectors to perform semantic line detection. On the other hand, these methods ignore the inherent properties of lines, which results in performance that is less than ideal. Lines have much more straightforward geometric properties compared to more complex objects, and as a result, they can have their parameters condensed down to just a few arguments. In this paper, we propose a one-shot end-to-end learning framework for line detection. Our goal is to better exploit the property of lines, so we incorporate the traditional Hough transform technique into deeply learned representations and do so using deeply learned representations. The Hough transform is used to translate deep representations into the parametric domain, which is where line detection takes place. Lines are parameterized with slopes and biases before the Hough transform is performed. To be more specific, we first assign features that have been aggregated to corresponding locations in the parametric domain, and then we assign features that have been aggregated along candidate lines on the feature map plane. As a consequence of this, the problem of locating semantic lines in the spatial domain is converted into the problem of locating individual points in the parametric domain, which makes the post-processing steps, specifically non-maximal suppression, more effective. In addition, our method makes it simple to extract contextual line features, which are essential for accurate line detection. These features are essential. In addition to the method that has been suggested, we also construct a large-scale dataset for the line detection task and design an evaluation metric that will assess the quality of the line detection. The results of our experiments on our proposed dataset as well as another public dataset demonstrate the advantages of our method in comparison to other state-of-the-art alternatives that have been used in the past.