benchmark/tables/release_tables/aide/analytic_baselines.md

AIDE analytic baselines — official test split (n=609, variant=protocol)

Protocol: driver rule first, grid floor ([0, 4, 9, 14, 19, 24, 29, 34, 39, 44]), HM = top 60 clips by mean displacement over 4 future transitions (threshold 72.05 px), PCK threshold t·max(W,H) (96 px @0.05). 95% CIs = clip-level bootstrap (10000 resamples). Δ vs ZV = paired per-clip difference (negative = better). Generated 2026-09-12T03:48:48Z by aide/eval/baselines_analytic.py; all numbers come from eval/harness.py output (metrics_test.json of each run).

run	All-MPJPE	d-nMPJPE	HM-MPJPE	HM d-nMPJPE	PCK@0.05	PCK@0.10	h=1	h=5	HM h=5	Δ vs ZV (All)	Δ vs ZV (HM)
zero_velocity	52.93 [48.88, 57.25]	2.403	139.23 [116.93, 167.68]	6.320	85.95	93.61	36.96	63.74	178.67	+0.00 [+0.00, +0.00]	+0.00 [+0.00, +0.00]
const_vel_a1.0	163.50 [143.17, 188.16]	7.422	342.29 [261.38, 473.58]	15.538	70.18	82.36	67.24	256.92	538.12	+110.57 [+91.37, +134.35]	+203.06 [+127.60, +333.05]
const_vel_a0.5	102.26 [91.33, 115.12]	4.642	228.63 [183.07, 295.19]	10.379	77.85	87.54	49.60	151.55	341.17	+49.33 [+39.94, +60.91]	+89.41 [+53.17, +151.68]
const_vel_a0.25	74.35 [67.70, 81.80]	3.375	177.42 [147.36, 216.12]	8.054	81.84	90.35	42.19	102.37	249.84	+21.42 [+16.89, +27.06]	+38.19 [+21.35, +67.16]
anti_velocity_0.1	52.53 [48.46, 56.84]	2.384	138.55 [116.22, 167.04]	6.289	85.83	93.73	37.00	63.02	177.22	-0.40 [-0.82, +0.13]	-0.68 [-1.67, +0.63]
mean_obs	55.94 [51.00, 61.55]	2.539	140.07 [116.80, 168.65]	6.358	84.49	93.26	46.68	62.38	163.87	+3.01 [-0.10, +7.04]	+0.84 [-6.28, +8.69]
per_part_copy	52.69 [48.61, 57.04]	2.392	138.57 [116.20, 166.93]	6.290	85.86	93.69	38.11	62.49	174.83	-0.24 [-0.92, +0.62]	-0.66 [-2.35, +1.10]
zv_shrink_val	51.54 [47.46, 55.97]	2.340	134.36 [111.80, 162.87]	6.099	85.88	93.91	38.37	60.54	167.45	-1.39 [-2.66, +0.26]	-4.86 [-7.67, -1.95]
mean_pose_shrink_a0.1	54.41 [50.54, 58.57]	2.470	135.66 [113.26, 164.22]	6.158	86.27	93.92	40.04	64.22	173.03	+1.48 [+0.96, +2.03]	-3.57 [-4.80, -2.41]
mean_pose_shrink_a0.2	59.31 [55.63, 63.35]	2.692	134.17 [111.59, 162.95]	6.090	84.92	94.08	46.85	67.91	168.88	+6.38 [+5.26, +7.57]	-5.06 [-7.44, -2.73]
mean_pose_shrink_a0.3	65.78 [62.16, 69.71]	2.986	134.09 [111.22, 162.97]	6.087	82.07	93.85	55.13	73.20	166.00	+12.85 [+11.14, +14.68]	-5.14 [-8.70, -1.62]
zv_shrink_h_val	51.21 [47.14, 55.63]	2.325	132.34 [109.76, 160.95]	6.007	86.13	93.91	36.65	60.89	163.95	-1.72 [-3.06, +0.02]	-6.89 [-10.11, -3.59]

Published (Driver-WM Table 1, typed from the paper into protocol.PAPER):

source	All-MPJPE	d-nMPJPE	HM-MPJPE	HM d-nMPJPE	PCK@0.05	HM h=5
paper zero-velocity	52.89	2.4	139.19	6.32	85.95	–
paper Driver-WM	71.47	3.24	138.03	6.27	71.66	155.82
paper static pooling (best HM)	–	–	134.56	6.11	–	–

Roadmap B0 rows: mean_pose_shrink_a{0.1,0.2,0.3} shrink the last observed pose towards the TRAIN-split mean pose (protocol frame 4, per joint; a fixed a priori); zv_shrink_h_val uses α_h = a·h/5 towards the observed mean with a fitted on val All. Trade-off table and Pareto figure: tradeoff_all_vs_hm.md, pareto_all_vs_hm.png.