Topological Inference for Occluded Infrastructure Networks

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 273 papers88 conceptsfrozen 2026-10-04

When we look at a satellite image of a city, our brains effortlessly connect the dots. If a road disappears under a thick canopy of trees or behind the shadow of a skyscraper, we don't assume the road has ended; we infer its path based on its trajectory and the surrounding context.

For traditional computer vision models, however, these "visual gaps" are catastrophic. Standard semantic segmentation—the process of labeling every pixel as "road" or "not road"—relies on direct visual evidence. When that evidence is obscured by shadows, vegetation, or clouds, the model produces fragmented, "broken" results. This fragmentation renders the data useless for navigation, where a single pixel-wide gap is the difference between a functional route and a dead end.

To solve this, researchers have moved beyond simple image processing toward topological inference: a set of techniques designed to "heal" infrastructure networks by treating them as graphs rather than just collections of pixels.

The Fragmentation Problem

The primary obstacle in automated mapping is occlusion. In high-resolution imagery, environmental factors like building shadows and tree canopies frequently mask the underlying geometry of roads and bridges Qing-Ling Shu et al. note that even advanced frequency-domain analysis struggles when the visual signal is completely masked.

This isn't just a problem for satellites. Ground-based urban auditing, as explored in the SIM framework, faces similar challenges from parked vehicles and dense vegetation. Even in the digital realm, Sebastien Callier and Hideo Saito show that extracting roads from printed maps is hindered by overlapping text labels and icons that create "holes" in the data.

From Pixels to Iterative Graph Construction

The most significant shift in this field is the move away from pixel-by-pixel segmentation in favor of iterative graph exploration. Instead of looking at the whole image at once, these models act like a search agent that "walks" along the road.

RoadTracer, developed by Songtao He and colleagues, uses a CNN to guide an agent that explores the network step-by-step. Because the agent is trained to look for the continuation of a path rather than just the presence of road-like pixels, it can maintain connectivity across junctions and small occlusions. Similarly, VecRoad uses iterative exploration with flexible step sizes to balance geometric precision with topological consistency.

Healing the Gaps with Global Constraints

If an agent gets lost in a dense forest canopy, how does it find the road on the other side? Modern frameworks use global properties to "repair" local failures.

The IGraSS framework by Abhijin Adiga and Madhav Marathe demonstrates this by enforcing graph-theoretic constraints. For example, if the model is mapping an irrigation canal, it "knows" that every segment must eventually connect back to a water source. By enforcing these global rules, the system can automatically repair noisy or incomplete ground truth data.

Other models, like LGNet, use auxiliary tasks to help the model "see" through shadows. By training the network to predict the general location of a road as a secondary goal, Jingtao Hu and Qi Wang show that the model learns to "fill in the gaps" where the primary segmentation fails.

Fusing Vision with Motion

Perhaps the most robust way to infer a hidden road is to watch how people use it. When visual evidence is absent, researchers turn to multi-modal fusion.

Tao Sun and Zonglin Di pioneered the use of crowdsourced GPS data as a supplementary "vision" layer. Even if a road is invisible under a tree canopy, the GPS traces of cars passing through it provide an unambiguous signal of its existence.

This idea is taken further in AutoGraph, where Ingmar Posner and Wolfram Burgard use vehicle "tracklets" as a self-supervised training signal. By aggregating the natural motion of traffic, the system learns to predict complex lane graphs without needing manual annotations, effectively using the behavior of agents to define the geometry of the space.

The Future: Generative and Sequential Modeling

The latest frontier treats map-making as a sequence generation task, similar to how a Large Language Model predicts the next word in a sentence. MARS (Map Auto-Regressor) treats map geometry as a "formal language," generating vectorized roads directly from imagery.

By moving from "classification" (is this pixel a road?) to "generation" (what is the most likely path of this road?), models like those proposed by Sotiris Anagnostidis and Aurelien Lucchi can use reinforcement learning and tree searches to navigate through occluded regions. These systems don't just see the road; they reason about where the road must be to maintain the integrity of the network.

Papers in this version

Showing 60 of 73 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 72 papers
  2. Version 1 · 2026-10-04 · 73 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_visual_occlusion_in_remote_sensing_0e408786,
  title = {visual occlusion in remote sensing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/visual-occlusion-in-remote-sensing/rdr_0e408786eb9f4563b326786ed56ec52c},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_topological_inference_for_occluded_infra_0e408786_v1,
  title = {Topological Inference for Occluded Infrastructure Networks},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/topological-inference-for-occluded-infrastructure-networks/rdr_0e408786eb9f4563b326786ed56ec52c/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}