Training-Free Building Change Detection with Example-Induced Discriminants

Zhi Lu, Bin Hong, Zhenming PengarXiv 2026 · arXiv

TL;DR

LIND (Linear INduced Discriminants) is a training-free framework for building change detection that turns a few annotated examples into a robust linear classifier using frozen foundation model features. By leveraging regional feature contrasts, it outperforms complex open-vocabulary methods in accuracy while being significantly faster and more memory-efficient.

What Problem Does This Paper Address?

Building Change Detection (BCD)—identifying where structures have been built or demolished between two satellite images—is critical for urban planning and disaster assessment. However, traditional supervised models require massive, region-specific datasets that are expensive to label and often fail to generalize to new geographic areas.

While "foundation models" like the Segment Anything Model (SAM) or CLIP offer a path forward, they present their own challenges. Text-based prompts often fail to capture the specific visual nuances of local architecture, and many existing training-free pipelines are "leaky," allowing errors to propagate through multiple stages of segmentation and temporal comparison. Furthermore, change annotations are inherently ambiguous: a mask tells you where a change occurred, but it doesn't explicitly label which of the two images contains the building and which contains the background.

Methods

The authors introduce LIND, which operates directly on the frozen feature space of a Vision Transformer (specifically DINOv3).

The core innovation lies in Discriminant Construction. Instead of training a new head, LIND uses a few example image pairs to "induce" a linear decision rule. It calculates the contrast between the annotated change region (the "interior") and its immediate surroundings (the "ring" and "exterior"). This contrast provides two vital pieces of information:

  1. Temporal disambiguation: It identifies which date actually contains the building by finding the date with the highest contrast against the surroundings.
  2. A building-separating direction: It defines a vector ($w$) in the high-dimensional feature space that distinguishes "building-ness" from "context-ness."
Figure 2. Overview of our method. Example-induced discriminants and locally matched query responses produce change scores. Point clouds are schematic.

During inference, query images are projected onto this induced direction to generate "building response" maps. The final change score is a combination of the difference in these responses and the raw feature dissimilarity between the two dates. To handle the "shaky camera" effect of satellite imagery, the method includes a local neighborhood search to compensate for registration errors.

Datasets

The authors evaluated LIND across several diverse benchmarks:

  • LEVIR-CD & WHU-CD: Standard high-resolution building change detection datasets.
  • S2Looking: A challenging dataset featuring side-looking satellite imagery.
  • SECOND: A multi-class dataset used to test LIND's extension to other categories like trees, water, and playgrounds.
  • DSIFN-CD: Used for evaluating general land-cover change.

Key Insights and Ideas

  • Incomplete Supervision is a Feature, Not a Bug: The paper’s "aha moment" is realizing that even though change maps are incomplete (they don't label unchanged buildings), the contrast they reveal is enough to build a general-purpose detector. A classifier "induced" from a few changed buildings can successfully recognize unchanged buildings in new scenes.
  • Frozen Features are Highly Reusable: By using DINOv3, the authors show that general-purpose visual representations are robust enough that a simple linear rule derived in one region can be applied to another without any backpropagation.
  • Simplicity Scales: By avoiding text-grounding and auxiliary segmentation models, LIND avoids the "semantic gap" and error propagation issues that plague open-vocabulary methods.

Key Results

LIND achieves state-of-the-art performance among training-free methods. In a 10-shot setting (using only 10 example pairs):

  • Accuracy: On LEVIR-CD, LIND reached 74.17% IoU, surpassing the strongest open-vocabulary baseline, ReA-OVCD (73.21%).
  • Efficiency: LIND is a lightweight powerhouse. It boasts a latency of just 0.36s per query and uses only 1.51 GiB of memory—roughly 2.5x faster and 2.6x more memory-efficient than OmniOVCD.
  • Qualitative Superiority: As seen in Figure 3, LIND is particularly effective at separating neighboring buildings and detecting small structures that other methods often miss or blur together.
Qualitative comparison showing LIND's ability to resolve dense building layouts compared to other training-free baselines.

Limitations and Future Work

  • Example Dependency: Like all few-shot methods, LIND's performance is tied to the quality and representativeness of the chosen examples. If the examples don't capture the diversity of the target scene, accuracy may dip.
  • Semantic Specificity: While LIND generalizes to classes like "water" or "trees," it relies on the target having a distinct visual contrast from its surroundings. Highly camouflaged targets might remain a challenge.
  • Future Directions: The authors suggest exploring automatic example selection and further investigating how these induced discriminants could be integrated into larger, multi-modal foundation models for even broader environmental monitoring tasks.

Summary updated

Abstract

Building change detection (BCD) relies on dense annotations that are costly to collect across regions. A few annotated examples offer a more economical way to specify the target, but change maps provide incomplete building supervision: they label only changed buildings and leave the building-containing date unspecified. Our key observation is that these annotations nevertheless reveal features that recognize buildings in other images, including unchanged ones. We present Linear INduced Discriminants (LIND), a training-free framework that turns regional feature contrasts into a reusable linear building classifier. Using a single frozen vision transformer, it unifies building recognition and temporal comparison in the same feature space. This approach requires neither text prompts nor an auxiliary segmentation model. Only one parameter is tuned on a single validation dataset and then kept fixed across datasets and shot settings. Extensive experiments demonstrate strong detection accuracy, cross-dataset transfer, and efficient inference, with extensions to other semantic classes and broader land-cover changes. Change annotations thus provide a compact visual specification for detection in new scenes. Code: https://github.com/lzcn/lind
BibTeX
@misc{lu2026training,
  author = {Zhi Lu and Bin Hong and Zhenming Peng},
  title = {{Training-Free Building Change Detection with Example-Induced Discriminants}},
  year = {2026},
  eprint = {2611.22964},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.22964},
}
Figure