VLM-Contact/configs/gate_a_oracle.yaml at main · BoPythonAI/VLM-Contact · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
experiment:
  id: GATE-A
  title: Oracle contact headroom
  scientific_status: pilot
  hypothesis: >-
    Under identical initialization and optimization budget, ground-truth
    contact edges improve interaction geometry without violating frozen
    pose, penetration, or reprojection non-inferiority margins.

dataset:
  name: CHI3D
  event_count_min: 50
  event_count_max: 100
  split_unit:
    - event_id
    - subject_pair
  first_pass_views_per_event: 1
  keep_all_views_of_event_in_same_split: true

paired_controls:
  identical_initializer: true
  identical_camera: true
  identical_base_losses: true
  identical_iteration_budget: true
  identical_random_seed: true
  only_contact_intervention_may_change: true

conditions:
  - id: O0
    contact_source: none
    description: Geometry-only optimization
  - id: O1
    contact_source: gt_top1
    description: One ground-truth contact edge
  - id: O2
    contact_source: gt_all
    description: All ground-truth contact edges
  - id: O3
    contact_source: oracle_best_edge_weight
    description: Diagnostic upper bound only
    eligible_for_method_selection: false

contact_weight_debug_grid:
  - 0.1
  - 0.3
  - 1.0
  - 3.0

optimizer_contact_intervention:
  loss_type: hhcmap
  release_reference_weight: 10.0
  pilot_frozen_weight: 0.1
  optimizer_max_iters: [1000, 1000]
  early_stopping:
    num_prev_steps: 20
    slope_tol: -0.0001
  implementation: ProsePose.llib.losses.contact.ContactMapLoss
  note: >-
    Do not use heuristic_02_chi3d.yaml's hhcdistmin for O0/O1/O2: that
    implementation ignores the supplied 75x75 contact map. The canonical
    hhcmap reference weight follows ProsePose flickr_fits.yaml. Weight 0.1
    was frozen for the noise pilot after the preregistered one-event debug
    grid because it preserved the strongest safety trade-off.

noise_pilot:
  event_count: 10
  optimizer_repeats: 3
  freeze_neutral_bands_before_main_pilot: true

# Frozen after observing zero optimizer-repeat noise in all 20 cells and before
# the 50-event main Gate-A pilot. These are round practical-resolution margins,
# not fitted to maximize the observed treatment effect.
frozen_outcome_thresholds:
  primary_contact_improvement_mm: 5.0
  contact_f1_change: 0.01
  safety_noninferiority:
    mpjpe_mm: 3.0
    pa_mpjpe_mm: 3.0
    pve_translation_aligned_mm: 3.0
    relative_translation_error_mm: 10.0
    penetration_max_mm: 10.0
    penetration_mean_mm: 3.0
    reprojection_error_px: 0.25
  helpful: >-
    GT-edge distance improves by more than 5 mm and every safety delta remains
    within its non-inferiority margin.
  harmful: >-
    Any safety delta exceeds its non-inferiority margin or GT-edge distance
    worsens by more than 5 mm.
  neutral: all remaining outcomes

metrics:
  primary:
    - contact_distance
    - contact_region_f1
  safety:
    - mpjpe
    - pa_mpjpe
    - pve
    - relative_translation_error
    - penetration
    - reprojection_error
  diagnostics:
    - optimizer_success
    - optimizer_iterations
    - runtime_seconds

outcome_label:
  classes:
    - helpful
    - neutral
    - harmful
  neutral_band_source: repeated_optimizer_noise
  non_inferiority_margin_source: preregistered_after_noise_pilot
  forbid_post_intervention_features_in_selector: true

gate:
  go:
    - paired_contact_improvement_cluster_bootstrap_ci_excludes_zero
    - all_frozen_safety_non_inferiority_constraints_hold
    - improvement_is_not_confined_to_a_tiny_event_subset
  no_go:
    - oracle_contact_has_no_stable_headroom
    - oracle_contact_systematically_harms_safety_metrics