Perception Under Occlusion
Robust unit classification in feather-dense environments
Abstract
Production processing environments present persistent occlusion, motion blur, and surface variation that degrade conventional classification pipelines. We describe a training curriculum combining 400,000 hours of line footage with synthetic occlusion augmentation, achieving robust unit classification and orientation estimation at full line speed.
The environment
Our facilities run dark: perception operates on infrared and structured-light arrays rather than visible-spectrum cameras. This removes flicker and glare, but introduces its own regime — low texture contrast, specular returns from process surfaces, and dense soft occluders in the intake stages. Units move at line speed under variable overlap, and the system must classify, orient, and act within an 11-millisecond decision budget.
Curriculum
We train against three data regimes: (1) 2.1 billion production-labeled frames sampled across facilities, seasons, and equipment generations; (2) synthetic occlusion augmentation, in which rendered occluders are composited over labeled frames with physically plausible depth ordering; and (3) hard-negative mining driven by production disagreement between model generations. Sensor dropout and calibration drift are simulated during training so the deployed model tolerates real-world mesh degradation.
Results
| Occlusion severity | Classification accuracy | Orientation accuracy |
|---|---|---|
| None (<10%) | 99.99% | 99.8% |
| Moderate (10–40%) | 99.96% | 99.3% |
| Heavy (40–70%) | 99.71% | 97.9% |
| Extreme (>70%) | 97.4% | 91.2% |
Units that fail classification at required confidence are routed to automated resolution. Extreme-occlusion performance remains an active research area.