The paper presents BorderNet-Natural, an extension of BorderNet, to handle complex occlusion patterns in natural images. It includes multi-scale oriented filters, an adaptive occlusion-aware attention module, and a hierarchical feature fusion mechanism. Experiments show state-of-the-art performance on occluded edge detection benchmarks.
Key findings
BorderNet-Natural (BorderNet-N) incorporates multi-scale oriented filters for boundary continuity.
An adaptive occlusion-aware attention module identifies and suppresses occluder regions.
Hierarchical feature fusion integrates low-level edge cues with high-level semantic features.
BorderNet-N achieves state-of-the-art performance on occluded edge detection benchmarks.
Limitations & open questions
The paper does not discuss the computational complexity of the proposed architecture.
The generalization of BorderNet-N to other types of occlusions beyond the tested datasets is not evaluated.