Abstract
Abstract
Edge detection is a fundamental task that aims to identify significant local changes in intensity within an image. It is widely used in various computer vision problems such as image segmentation, depth estimation, object detection, and optical flow estimation. This thesis revisits the edge detection problem by addressing three main challenges: (P1) the commonly used loss functions for addressing class imbalance, (P2) the exploitation of label uncertainty due to multiple annotations for a single RGB image, and (P3) the generation of one-pixel-wide edge maps in an end-to-end trainable manner to overcome the common issue of thick edge predictions in existing methods.
To jointly address P1 and P2, we introduce the use of ranking-based loss functions as a fundamentally new approach to edge detection. RankED has two main components: A ranking loss and a sorting loss. The Ranking Loss mitigates the severe imbalance between edge and background pixels whereas the Sorting Loss prioritizes high-confidence edges over low-confidence ones based on computed label uncertainty. We experimentally show that RankED consistently outperforms existing approaches on three widely-used edge detection datasets, NYUDv2, BSDS500, and Multi-Cue.
We further extend RankED by integrating hyperbolic embedding, resulting in RankHED, which achieves new state-of-the-art performance on the NYUDv2, BSDS500, and Multi-Cue datasets across ODS, OIS, and AP metrics.
Finally, we propose a new matching-based supervision method, MatchED, which can be integrated into any edge detection pipeline to address P3. We demonstrate that MatchED not only yields approximately 20-point improvements for SOTA edge detection methods in ODS, OIS, and AP under crispness-emphasized evaluation protocols (CEval), but also outperforms NMS and skeleton-based thinning algorithms under standard evaluation (SEval)