Line Detection
Our line detection algorithm is composed of two stages. Line region proposals in the form of rotated bounding boxes are first generated; then, these are filtered using a binary classifier.
For the region proposal, we use the Hough transform for lines. To reduce the latency of this step, the input image is first resized to a lower resolution. We observed that downscaling by a factor of 4 i.e. to a resolution of 160x120px, a good trade-off between runtime performance and accuracy can be achieved.
A rotated bounding box is extracted around each line region proposal and passed to a Random Forest line segment classifier. The perfomance of this type of classifiers usually depends on the discriminative power of the input features, with more sophisticated features usually achieving better results. That being said, in our experiments we observed that simple features, such as the difference in intensity between pixels within the line proposal patch are discriminative enough.