Occlusion can be detected without training or labels by exploiting the disagreement between apparent height and monocular depth—two signals a standard vision pipeline already produces. This matters because occlusion annotations are rarely available in real deployment settings.
This paper proposes a simple, training-free method to detect when pedestrians are occluded in images by comparing two distance estimates a camera already computes: apparent height (from bounding boxes) and monocular depth. Occlusion pulls depth estimates toward the nearest surface while leaving apparent height unchanged, so their disagreement signals occlusion.