Dhi Labs: Fixed-Camera 3D (A3) demo

No LiDAR, no stereo rig, no site visit. Self-calibrates height, tilt and focal length purely from the people already walking through the scene, then turns every detection box a downstream detector produces into a real-world position, distance, height, speed, separation, area and density.

What is measured, what is not.

The headline table and the full grid below are measured this session with numpy==2.5.1, scipy==1.18.0, seed 0, 80 calibration walkers and 60 evaluation points per cell.
Real annotated CCTV footage has not been run through this pipeline yet. Every number here is synthetic ground truth. The interface for real-footage evaluation is written out in docs/EXTERNAL_EVAL.md, not run yet (board item AC1).

Headline sweep (DEFAULT_SWEEP, three configurations)

These three configurations are the ones committed as evidence in the repository itself (src/fixedcam3d/bench.py, DEFAULT_SWEEP). They are the numbers the product README states, reproduced digit for digit by running the command below in this generation session.

camera (truth)	position RMSE	position p90	speed MAE	height MAE	recovered tilt error	sensitivity (m/0.5°)
3 m, 20 deg tilt (hard case)	4.30 m	5.77 m	0.353 m/s	0.038 m	3.14 deg	1.72
5 m, 30 deg tilt	0.16 m	0.23 m	0.083 m/s	0.056 m	0.43 deg	0.72
8 m, 45 deg tilt	0.19 m	0.22 m	0.076 m/s	0.078 m	0.02 deg	0.2

Command: fixedcam3d bench --seed 0 (python -m fixedcam3d.cli bench --seed 0), reading src/fixedcam3d/bench.py. Full suite this session: python -m pytest tests/ -q printed 51 passed in 13.37s (tests/test_bench.py, tests/test_cli.py, tests/test_depth_fusion.py, tests/test_geometry.py, tests/test_metric.py). The raw grid backing this page (bench_results.json) is committed alongside this page in the Space itself, not just claimed.

Extended height / tilt grid, all 42 configurations

This is every cell behind the interactive grid above, as plain numbers. Rows shaded and bolded are the ones the calibrator's own conditioning probe flags as ill-conditioned (sensitivity_m_per_half_deg > 1.0, the same threshold calibration.py uses internally): 11 of the 42 cells cross that line, concentrated at low height paired with a shallow tilt.

height	tilt	focal	position RMSE (m)	position p90 (m)	speed MAE (m/s)	height MAE (m)	sensitivity (m/0.5°)	residual
3 m	15 deg	1200 px	6.681	9.094	0.440	0.032	2.03	0.0366
3 m	20 deg	1200 px	4.304	5.767	0.353	0.038	1.72	0.0366
3 m	25 deg	1200 px	3.040	3.992	0.298	0.044	1.57	0.0366
3 m	30 deg	1200 px	1.278	1.631	0.213	0.066	0.95	0.0399
3 m	35 deg	1200 px	1.172	1.317	0.234	0.065	0.43	0.0387
3 m	40 deg	1200 px	1.213	1.316	0.165	0.181	0.15	0.0380
3 m	45 deg	1200 px	6.031	6.397	n/a	0.057	0.31	0.0357
4 m	15 deg	1300 px	0.571	0.796	0.209	0.035	1.08	0.0365
4 m	20 deg	1300 px	1.688	2.205	0.277	0.046	1.17	0.0409
4 m	25 deg	1300 px	2.008	2.516	0.268	0.049	1.11	0.0392
4 m	30 deg	1300 px	1.070	1.296	0.113	0.056	0.76	0.0395
4 m	35 deg	1300 px	0.215	0.279	0.094	0.094	0.33	0.0450
4 m	40 deg	1300 px	0.640	0.684	0.114	0.124	0.16	0.0398
4 m	45 deg	1300 px	0.052	0.073	0.123	0.135	0.14	0.0381
5 m	15 deg	1400 px	2.094	2.585	0.100	0.034	0.81	0.0352
5 m	20 deg	1400 px	0.211	0.283	0.110	0.039	0.91	0.0366
5 m	25 deg	1400 px	0.437	0.587	0.157	0.052	0.88	0.0412
5 m	30 deg	1400 px	0.165	0.235	0.083	0.056	0.72	0.0390
5 m	35 deg	1400 px	0.513	0.587	0.131	0.073	0.41	0.0377
5 m	40 deg	1400 px	1.644	1.800	0.069	0.067	0.25	0.0361
5 m	45 deg	1400 px	0.253	0.280	0.039	0.120	0.14	0.0417
6 m	15 deg	1500 px	47.410	61.124	1.740	0.024	5.94	0.0397
6 m	20 deg	1500 px	1.264	1.478	0.077	0.037	0.75	0.0353
6 m	25 deg	1500 px	0.101	0.145	0.097	0.043	0.79	0.0366
6 m	30 deg	1500 px	0.394	0.515	0.163	0.048	0.73	0.0403
6 m	35 deg	1500 px	0.454	0.541	0.171	0.062	0.36	0.0396
6 m	40 deg	1500 px	0.581	0.628	0.071	0.084	0.24	0.0391
6 m	45 deg	1500 px	0.494	0.542	0.091	0.114	0.14	0.0398
7 m	15 deg	1600 px	60.594	74.534	1.231	0.026	7.25	0.0369
7 m	20 deg	1600 px	12.900	15.969	0.461	0.025	1.42	0.0397
7 m	25 deg	1600 px	0.838	0.972	0.067	0.040	0.69	0.0354
7 m	30 deg	1600 px	0.109	0.167	0.091	0.046	0.65	0.0361
7 m	35 deg	1600 px	1.457	1.570	0.073	0.076	0.3	0.0396
7 m	40 deg	1600 px	0.143	0.181	0.083	0.092	0.22	0.0391
7 m	45 deg	1600 px	0.461	0.496	0.042	0.084	0.18	0.0394
8 m	15 deg	1700 px	6.555	8.186	0.166	0.025	1.18	0.0328
8 m	20 deg	1700 px	31.846	38.362	0.554	0.028	2.82	0.0368
8 m	25 deg	1700 px	7.457	8.942	0.270	0.027	0.94	0.0396
8 m	30 deg	1700 px	0.593	0.681	0.060	0.044	0.63	0.0355
8 m	35 deg	1700 px	1.148	1.267	0.081	0.056	0.34	0.0430
8 m	40 deg	1700 px	1.424	1.639	0.074	0.083	0.22	0.0366
8 m	45 deg	1700 px	0.190	0.219	0.076	0.078	0.2	0.0392

Grid: fixedcam3d.bench.run_bench(), the same function DEFAULT_SWEEP calls, swept over height 3 to 8 m (step 1 m) by tilt 15 to 45 degrees (step 5 degrees), seed 0, 80 calibration walkers and 60 evaluation points per cell, focal length scaled with height the same way the shipped sweep does (focal_px = 900 + 100 * height_m, confirmed against all three DEFAULT_SWEEP cameras). This 42-cell grid is not one of the product repo's committed evidence files (the repo ships only the 3-camera DEFAULT_SWEEP); it is a disclosed extension built from the identical run_bench code path, computed fresh in this generation session and committed in bench_results.json in this Space.

Fixed-Camera 3D: metric 3D from one camera you already have

Headline sweep (DEFAULT_SWEEP, three configurations)

Extended height / tilt grid, all 42 configurations

Honest limitations