Paper: Occlusion for Free: Detecting Occlusion from the Disagreement Between Apparent Height and Monocular Depth Field: Computer Vision Authors: Chibueze Praise Onyenweaku, Toyibat Adele, Sikirulahi Opeyemi Abdulkareem, Obed Honour Eje Published: 2026-11-04 arXiv: 2611.06899
Summary
Knowing that a pedestrian is occluded is useful to a robot even when the detector cannot recover the hidden pixels, but occlusion estimators are normally trained on occlusion annotations that deployment settings rarely have. This paper shows that occlusion can be read off two distance estimates that a camera-only pipeline already computes.
Key Idea
- Apparent height is largely unchanged by occlusion, because the detected box still spans the whole person.
- Monocular metric depth is pulled toward the occluder, because the nearest surface inside the box is what a depth map reports.
- Only the ordering of the two estimates matters, so no focal length, stature, or depth scale is needed—provided the depth model's error is close to a pure gain, a condition that can be checked without labels.
- KITTI: heavy occlusion detection AUC 0.909 (0.888 over all annotated pedestrians, including ones the detector misses), with no training and no occlusion labels.
- The score is not a distance proxy: within fixed distance bins, AUC is 0.894–0.925, while true distance alone is near chance.
- CARLA synthetic scenes (with measured visibility): AUC 0.819, versus 0.644 for distance alone.
- arXiv: https://arxiv.org/abs/2611.06899
Results
Context and Motivation
The authors measure how three standard architecture choices respond to occlusion—depth fusion, occlusion-simulation augmentation, and attention—and find that none of the tested configurations improved out-of-distribution detection. The method emerged after observing that a detector trained on German road data lost half its precision on self-annotated street imagery from Lagos, Nigeria. In environments not represented by benchmarks, a signal that requires no local annotations is especially valuable.