File size: 11,903 Bytes
540e67a
 
 
 
 
 
 
 
 
a8124a8
540e67a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
df8f96e
540e67a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a8124a8
540e67a
 
df8f96e
 
540e67a
df8f96e
540e67a
 
 
df8f96e
 
540e67a
df8f96e
540e67a
 
 
 
 
 
 
df8f96e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
540e67a
 
 
 
 
 
 
 
 
 
 
df8f96e
540e67a
 
 
 
 
 
 
df8f96e
 
 
 
540e67a
 
 
 
 
 
 
 
df8f96e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
540e67a
 
 
 
 
 
a8124a8
540e67a
 
df8f96e
 
 
540e67a
 
 
df8f96e
 
 
540e67a
 
 
 
 
 
 
df8f96e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
540e67a
 
 
 
 
 
 
 
a8124a8
540e67a
 
 
df8f96e
 
 
540e67a
 
 
df8f96e
 
 
540e67a
 
 
 
 
 
 
df8f96e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
540e67a
 
 
 
 
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
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
{
  "source": {
    "shared_windows": "results/episode_task_suite/shared_windows.npz",
    "windows_csv": "results/episode_task_suite/windows.csv",
    "feature_manifest": "results/episode_task_suite/feature_manifest.json"
  },
  "dataset_scope": {
    "sample_episode_count": 1,
    "num_windows": 1161,
    "feature_dim": 8546,
    "first_start_frame": 0,
    "last_end_frame": 5819,
    "warning": "Single public sample episode; these extension probes validate task design and pipeline mechanics, not cross-episode generalization."
  },
  "baselines": {
    "minimal": "Ridge classifiers/regressors/projections plus cosine retrieval on the committed feature tensor.",
    "neural_mlp": "Small one-hidden-layer PyTorch MLP heads using the same inputs, targets, chronological split, and evaluator."
  },
  "run_config": {
    "train_fraction": 0.7,
    "ridge_l2": 10.0,
    "seed": 7,
    "future_windows": 4,
    "neural_epochs": 25,
    "neural_hidden_dim": 128,
    "neural_batch_size": 128,
    "skip_neural": false
  },
  "task_specs": {
    "body_motion_intensity": {
      "direction": "A",
      "direction_name": "Human Modeling & Motion Understanding",
      "name": "Body and Hand Motion Intensity",
      "family": "classification",
      "case_study": "A window with a fast reach or pour should be classified as high motion; a steady holding window should be low motion.",
      "input": "Current non-mocap feature blocks: video, audio, depth, camera pose/rotation, IMU, SLAM, calibration, and language context.",
      "middle_process": "Compute the target from hand/body joint changes between neighboring windows, hide the mocap blocks from the input, then classify high versus low motion using the train-set median as the threshold.",
      "output": "Binary label: high_motion or low_motion.",
      "minimal_baseline": "Ridge classifier on standardized non-mocap features.",
      "neural_baseline": "One-hidden-layer MLP binary classifier on the same input features.",
      "metric_name": "macro-F1",
      "metric_key": "macro_f1",
      "metric_direction": "higher",
      "current_limit": "This is a motion-energy proxy, not a SMPL/MANO body model or a generative motion prior."
    },
    "multi_view_consistency_retrieval": {
      "direction": "B",
      "direction_name": "3D/4D Reconstruction & Neural Rendering",
      "name": "Multi-View Consistency Retrieval",
      "family": "retrieval",
      "case_study": "Given the fisheye camera features for a pouring moment, retrieve the synchronized stereo-left view from the same time window.",
      "input": "Query side: fisheye_cam0 video feature block. Candidate side: stereo_left video feature block from held-out windows.",
      "middle_process": "Learn a projection from one camera-view feature space into another, then rank held-out candidate windows by cosine similarity.",
      "output": "Ranked candidate windows; the correct synchronized view should rank near the top.",
      "minimal_baseline": "Ridge projection followed by cosine nearest-neighbor retrieval.",
      "neural_baseline": "One-hidden-layer MLP projection followed by the same cosine retrieval evaluator.",
      "metric_name": "MRR",
      "metric_key": "mrr",
      "metric_direction": "higher",
      "current_limit": "This checks calibrated multi-view signal, but it is still feature retrieval, not NeRF, Gaussian Splatting, or novel-view synthesis."
    },
    "action_phase_progress": {
      "direction": "C",
      "direction_name": "Egocentric Vision & Interaction",
      "name": "Action Phase Progress Estimation",
      "family": "regression",
      "case_study": "Inside a Pour coffee action segment, estimate whether the current window is near the beginning, middle, or end of that action.",
      "input": "Current non-caption multimodal feature vector, so the label text cannot be copied directly from the language block.",
      "middle_process": "Convert contiguous action-label runs into a normalized 0-to-1 progress target, train on earlier windows, and regress progress for later windows.",
      "output": "A scalar progress value between 0.0 and 1.0 for the current action segment.",
      "minimal_baseline": "Ridge regressor on standardized non-caption features.",
      "neural_baseline": "One-hidden-layer MLP regressor on the same input features.",
      "metric_name": "MAE",
      "metric_key": "mae",
      "metric_direction": "lower",
      "current_limit": "This is an action-structure probe inside one episode, not a general intent model across homes, people, or tasks."
    },
    "ego_motion_forecast": {
      "direction": "D",
      "direction_name": "Scene Reconstruction & World Modeling",
      "name": "Short-Horizon Ego-Motion Forecasting",
      "family": "forecast",
      "case_study": "From the current sensors, predict how the camera translation will change over the next 20 frames while the wearer moves through the scene.",
      "input": "Current multimodal features excluding the camera-translation block and caption text.",
      "middle_process": "Build a future target from camera-translation difference at a four-window horizon, then regress that future ego-motion delta from current sensors.",
      "output": "A future camera-translation delta vector.",
      "minimal_baseline": "Ridge regressor with a 20-frame forecast horizon.",
      "neural_baseline": "One-hidden-layer MLP regressor with the same horizon and split.",
      "metric_name": "MAE",
      "metric_key": "mae",
      "metric_direction": "lower",
      "current_limit": "This is a compact world-model proxy; it does not build a persistent map, scene graph, or object permanence model."
    }
  },
  "tasks": {
    "body_motion_intensity": {
      "train_windows": 812,
      "test_windows": 348,
      "target_threshold_train_median": 0.476467490196228,
      "input_dim": 6425,
      "target_source": "hand/body joint delta between neighboring windows",
      "minimal": {
        "accuracy": 0.7758620689655172,
        "macro_f1": 0.7658385093167701,
        "positive_rate_true": 0.35919540229885055,
        "positive_rate_pred": 0.4339080459770115,
        "num_test": 348
      },
      "neural_mlp": {
        "accuracy": 0.8304597701149425,
        "macro_f1": 0.8254423029509534,
        "positive_rate_true": 0.35919540229885055,
        "positive_rate_pred": 0.47126436781609193,
        "num_test": 348
      },
      "neural_training": {
        "available": true,
        "epochs": 25,
        "hidden_dim": 128,
        "loss_history": [
          0.37322977610996794,
          0.22245765099384515,
          0.1382973729242832,
          0.10363741681493562,
          0.0795709453523159,
          0.06539858697817244,
          0.055655122610735776,
          0.043255199022187385,
          0.03558319240001035,
          0.031631215764530776,
          0.029465350402711796,
          0.024383640274625695,
          0.02020622924740972,
          0.016222001351599624,
          0.018758724778523587,
          0.013199950316049196,
          0.014794624612432689,
          0.01013119441452505,
          0.009688532855040554,
          0.008956241283767622,
          0.006733611014469761,
          0.006677041435843618,
          0.0067647325489761795,
          0.005346325666556511,
          0.004052691048084588
        ]
      }
    },
    "multi_view_consistency_retrieval": {
      "train_windows": 813,
      "test_windows": 348,
      "query_block": "video_fisheye_cam0",
      "target_block": "video_stereo_left",
      "query_dim": 686,
      "target_dim": 686,
      "minimal": {
        "mrr": 0.5533982515335083,
        "top1": 0.41954022988505746,
        "top5": 0.7068965517241379,
        "top10": 0.8304597701149425,
        "median_rank": 2.0,
        "num_test": 348
      },
      "neural_mlp": {
        "mrr": 0.34691643714904785,
        "top1": 0.23275862068965517,
        "top5": 0.46264367816091956,
        "top10": 0.5890804597701149,
        "median_rank": 7.0,
        "num_test": 348
      },
      "neural_training": {
        "available": true,
        "epochs": 25,
        "hidden_dim": 128,
        "loss_history": [
          0.9800718805740094,
          0.8296866191855803,
          0.7029420470986126,
          0.6089339927846948,
          0.5426930648814268,
          0.49323386093200644,
          0.45315542230600214,
          0.4240272395578551,
          0.3964498403400806,
          0.37567753094588696,
          0.3599070675332021,
          0.3417643405048023,
          0.32952829051721577,
          0.31516501450450657,
          0.3070896395824639,
          0.29752101269888553,
          0.287490411878072,
          0.2791558311654193,
          0.2707079971921693,
          0.2669465311998811,
          0.2603630047442728,
          0.2501040017656148,
          0.24714160980920216,
          0.24146720613060843,
          0.23866056472173036
        ]
      }
    },
    "action_phase_progress": {
      "train_windows": 813,
      "test_windows": 348,
      "input_dim": 7650,
      "target_source": "normalized position inside contiguous action-label runs",
      "minimal": {
        "mse": 0.16943013668060303,
        "mae": 0.32674381136894226,
        "r2": -1.0236103208433347,
        "num_test": 348
      },
      "neural_mlp": {
        "mse": 0.14226463437080383,
        "mae": 0.301545649766922,
        "r2": -0.6991557278041094,
        "num_test": 348
      },
      "neural_training": {
        "available": true,
        "epochs": 25,
        "hidden_dim": 128,
        "loss_history": [
          3.236165899632162,
          1.2893786148831414,
          0.8036823107014429,
          0.5113777590120557,
          0.42643894586909153,
          0.37627028047965166,
          0.30304253713524504,
          0.25041163572526065,
          0.20774717810409096,
          0.18116216590630319,
          0.16409150619165191,
          0.14362229159397394,
          0.1277421933001991,
          0.1232111468672899,
          0.11924667946013724,
          0.11960234325310401,
          0.09951645682988572,
          0.08239271907825459,
          0.08965909919102401,
          0.07973466079503408,
          0.07939992471154765,
          0.07085797808340408,
          0.08331163055269188,
          0.068286436959313,
          0.07575550977814241
        ]
      }
    },
    "ego_motion_forecast": {
      "train_windows": 810,
      "test_windows": 347,
      "forecast_horizon_windows": 4,
      "forecast_horizon_frames": 20,
      "input_dim": 7629,
      "target_dim": 21,
      "target_source": "future minus current camera_translation feature block",
      "minimal": {
        "mse": 3.3189830780029297,
        "mae": 0.16999860107898712,
        "r2": -5674.96718626448,
        "num_test": 347
      },
      "neural_mlp": {
        "mse": 0.547074019908905,
        "mae": 0.0833469107747078,
        "r2": -934.5800418396838,
        "num_test": 347
      },
      "neural_training": {
        "available": true,
        "epochs": 25,
        "hidden_dim": 128,
        "loss_history": [
          1.034793225188314,
          0.6070329941349265,
          0.459870958622591,
          0.3616479166495947,
          0.2902924293353234,
          0.23913239262722158,
          0.1907091121982645,
          0.17297036942140556,
          0.15034288657300265,
          0.13807891327657817,
          0.13084740807980666,
          0.12387588925567675,
          0.12085371225336451,
          0.11271689225126195,
          0.10728766140010622,
          0.10044601888936243,
          0.09365091108613544,
          0.09090288755150489,
          0.08592776887946658,
          0.0805281208805096,
          0.08014915316929051,
          0.07900124887625376,
          0.07897219589830917,
          0.07575527565714754,
          0.0774567014273302
        ]
      }
    }
  }
}