Corrects a raw detection count for what the camera cannot see, instead of reporting the visible-only number as if it were the true one.
Closed source, inquireAmodal counting combines box-geometry visibility estimation, inference-aligned detectability calibration, Horvitz-Thompson interval counting with conformal calibration, and track-based occlusion persistence into one visibility-corrected crowd count.
The project has gone through multiple real-data iterations, each one documented as surfacing and then partially fixing a real methodological problem, rather than jumping straight to a clean final number.
| Crowding | Naive MAE | Corrected MAE | 90% CI coverage |
|---|---|---|---|
| 0.0 | 2.8 | 2.4 | 0.90 |
| 0.3 | 2.6 | 1.97 | 0.90 |
| 0.6 | 2.85 | 2.34 | 0.975 |
| 0.8 | 3.73 | 2.46 | 0.975 |
The original one-dimensional visibility-only detectability curve came back nearly flat (56 to 60 percent recall). A multivariate detectability model fit better but still did not beat naive counting, because real detector error mixes misses and false positives, and Horvitz-Thompson reweighting can only push counts up. A follow-up two-sided estimator is the first approach in this trail to beat naive counting on real data in aggregate: MAE 11.767 to 11.031, with 90.67 percent empirical coverage against a 90 percent target, with one disclosed exception, the single densest crowd scenes, where naive counting still wins and the estimator's own coverage flag correctly signals extrapolation.
An outsider can download the published benchmark dataset and try the demo Space to see the synthetic calibration behavior. The CrowdHuman real-data result is reported from the project's own files and is not independently re-runnable from outside, since the source code is not public.