Challenges:

• Most existing vector extraction works focus on a specific target and are fragile to category variety;

• Especially Line- and polygon-shape objects are very different in topological structure.

Contributions:

• We propose an innovative class-agnostic model TopDiG,to directly extract topological directional graphs from remote sensing images regardless of target types.

• We design topology-concentrated node detector to extract nodes and corresponding features for both line-and polygon-shape targets.

• We conduct dynamic graph supervision strategy to facilitate inference in practice .

TopDiG can extract vector topology of both line- and polygon-shape objects

TopDiG formulates diverse vector topological structures as directional graphs and works by sequentially predicting target nodes and their connections.

Topology-Concentrated Node detector(TCND):

• In satellite and aerial images, line shape objects can be seen as edges

• TCND minimizes sematic contexts and concentrates on low-level geometric features to achieve category independency.

Visual comparison of attentive maps for a few methods

Dynamic Graph Supervision(DGS):

• Dynamically generates adjacency matrix labels according to real-time unordered predict nodes in each training.

Visual comparison of the adjacency matrix label generation by previous works and DGS

Visual comparison:

• TopDiG performs better in topology quality and completeness, especially in concave/hierarchical polygon contours and parallel/crossing line object centerlines.

Quantitative comparison:

• TopDiG obtains competitive scores in pixel-wise metrics

• TopDiG achieve state-of-the-art performance in topology-wise metrics