The dense_direction is a framework for dense direction estimation of linear objects in images.
It is built on top of popular OpenMMLab's libraries (mmengine, mmcv, and mmseg), and utilizes loss-based
algorithmic weak-supervision to learn the direction estimation of linear objects from semantic segmentation maps.
- Loss-Based Weak-Supervision: My approach leverages loss functions to guide the learning process, reducing the need for explicit labeling. There is no need for direction labels, only widely available semantic segmentation maps are needed.
- Integration with Semantic Segmentation: The framework seamlessly integrates with semantic segmentation techniques, providing the means to train it alongside direction estimation.
- Extension of OpenMMLab's frameworks: By building upon MMSegmentation, we can tap into its established library and expertise in semantic segmentation.
It is a dense task that estimates the local direction for every pixel belonging to a linear structure (e.g., cracks, roads, fibers, vessels). Unlike edge detection or segmentation alone, it provides a continuous direction field that describes “which way the structure runs” at each location.
The plots show a colored overlay that encodes the estimated per‑pixel direction for pixels classified as concrete crack, blood vessel, and road in the example images from Concrete-Cracks, CHASE-DB1 and Ottawa-Roads datasets.
- Directions are estimated in the range [0, π), i.e., direction is modulo 180° (0° and 180° represent the same direction along a line).
- Angles increase counter‑clockwise from the image x‑axis (horizontal to the right).
- This convention is common for undirected linear structures; it avoids ambiguity between opposite directions along the same line.
- Enables downstream tasks that need local direction, such as topology analysis, path following, and measuring directional consistency or curvature along linear objects.
- Complements semantic segmentation by adding geometric context (direction) without requiring explicit direction labels.