Scikit-learn
human-activity-recognition
wearable
wrist
time-series
cpu
scikit-learn
Zipeng365 commited on
Commit
3a372ea
·
verified ·
1 Parent(s): 8910582

Refresh bilingual research videos with MiMo narration and synchronized subtitles

Browse files
.gitattributes CHANGED
@@ -37,3 +37,11 @@ videos/WISP_en_system_voice.mp4 filter=lfs diff=lfs merge=lfs -text
37
  videos/WISP_en_cover.png filter=lfs diff=lfs merge=lfs -text
38
  videos/WISP_zh_system_voice.mp4 filter=lfs diff=lfs merge=lfs -text
39
  videos/WISP_zh_cover.png filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
37
  videos/WISP_en_cover.png filter=lfs diff=lfs merge=lfs -text
38
  videos/WISP_zh_system_voice.mp4 filter=lfs diff=lfs merge=lfs -text
39
  videos/WISP_zh_cover.png filter=lfs diff=lfs merge=lfs -text
40
+ videos/archive/20261023/WISP_en_cover.png filter=lfs diff=lfs merge=lfs -text
41
+ videos/archive/20261023/WISP_en_system_voice.mp4 filter=lfs diff=lfs merge=lfs -text
42
+ videos/archive/20261023/WISP_zh_cover.png filter=lfs diff=lfs merge=lfs -text
43
+ videos/archive/20261023/WISP_zh_system_voice.mp4 filter=lfs diff=lfs merge=lfs -text
44
+ videos/WISP_zh_mimo_voice.mp4 filter=lfs diff=lfs merge=lfs -text
45
+ videos/WISP_zh_mimo_voice.wav filter=lfs diff=lfs merge=lfs -text
46
+ videos/WISP_en_mimo_voice.mp4 filter=lfs diff=lfs merge=lfs -text
47
+ videos/WISP_en_mimo_voice.wav filter=lfs diff=lfs merge=lfs -text
README.md CHANGED
@@ -30,22 +30,26 @@ Benchmark data remain in [WristHARBench on Hugging Face](https://huggingface.co/
30
 
31
  Two independently narrated animations explain WristHARBench, how WISP representations read wrist-motion signals, how complementary information is combined, and the experimental findings.
32
 
33
- | Language | Video | Subtitles | Script |
34
- | --- | --- | --- | --- |
35
- | English · 3:16 | [Watch / download MP4](https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_en_system_voice.mp4) | [SRT](videos/WISP_en.srt) · [VTT](videos/WISP_en.vtt) | [English script](videos/WISP_en_script.txt) |
36
- | 中文 · 3:12 | [观看 / 下载 MP4](https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_zh_system_voice.mp4) | [SRT](videos/WISP_zh.srt) · [VTT](videos/WISP_zh.vtt) | [中文脚本](videos/WISP_zh_script.txt) |
 
 
37
 
38
  ### English
39
 
40
- <video controls preload="none" width="100%" poster="https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_en_cover.png" src="https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_en_system_voice.mp4"></video>
41
 
42
  ### 中文
43
 
44
- <video controls preload="none" width="100%" poster="https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_zh_cover.png" src="https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_zh_system_voice.mp4"></video>
 
 
45
 
46
- Produced on 18 October 2026. Both videos use synthetic system narration and burned-in subtitles. Signal landscapes include illustrative synthetic inputs, distinguished from experimental results in the videos. The animations explain the public results checked at production time; they are not a live results dashboard.
47
 
48
- [Paper and manuscript (Pure)](https://research.birmingham.ac.uk/en/publications/wisp-composable-representations-for-wrist-worn-activity-recogniti/) · [WristHARBench](https://huggingface.co/datasets/Zipeng365/WristHARBench)
49
 
50
  ## Install
51
 
 
30
 
31
  Two independently narrated animations explain WristHARBench, how WISP representations read wrist-motion signals, how complementary information is combined, and the experimental findings.
32
 
33
+ **Narration updated on 28 October 2026:** Chinese uses MiMo's 茉莉 voice; English uses Chloe. Both are AI-generated voices. Subtitles and scene timing have been rebuilt for the new narration.
34
+
35
+ | Language | Video | Audio | Subtitles | Script |
36
+ | --- | --- | --- | --- | --- |
37
+ | English · 3:46 | [Watch / 下载 MP4](https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_en_mimo_voice.mp4) | [WAV](https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_en_mimo_voice.wav) | [SRT](videos/WISP_en.srt) · [VTT](videos/WISP_en.vtt) | [English script](videos/WISP_en_script.txt) |
38
+ | 中文 · 3:47 | [Watch / 下载 MP4](https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_zh_mimo_voice.mp4) | [WAV](https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_zh_mimo_voice.wav) | [SRT](videos/WISP_zh.srt) · [VTT](videos/WISP_zh.vtt) | [中文脚本](videos/WISP_zh_script.txt) |
39
 
40
  ### English
41
 
42
+ <video controls preload="none" width="100%" poster="https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_en_cover.png" src="https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_en_mimo_voice.mp4"></video>
43
 
44
  ### 中文
45
 
46
+ <video controls preload="none" width="100%" poster="https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_zh_cover.png" src="https://huggingface.co/Zipeng365/WISP/resolve/main/videos/WISP_zh_mimo_voice.mp4"></video>
47
+
48
+ The research story and graphics are retained from the 18 October 2026 production. Signal landscapes include illustrative synthetic inputs, distinguished from experimental results in the videos. This update changes narration and subtitle timing; it does not introduce new experimental results.
49
 
50
+ [Reproducible project](videos/WISP_MiMo_reproducible_project.zip) · [Reproduction instructions and sources](videos/REPRODUCE.md) · [Paper (Pure)](https://research.birmingham.ac.uk/en/publications/wisp-composable-representations-for-wrist-worn-activity-recogniti/)
51
 
52
+ [Original system-voice edition and subtitles](https://huggingface.co/Zipeng365/WISP/tree/main/videos/archive/20261023). Previously shared `system_voice.mp4` URLs now serve the refreshed editions.
53
 
54
  ## Install
55
 
videos/REPRODUCE.md ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # WISP bilingual narration refresh — 28 October 2026
2
+
3
+ This release preserves the research story and graphics from the 18 October 2026 animation and replaces macOS system speech with MiMo-V2.5-TTS narration (Chinese: 茉莉; English: Chloe). The voices are AI-generated. Pronunciation guidance, the exact narration text, and generated source audio are included.
4
+
5
+ The original release is preserved separately. Public files with the historic `system_voice.mp4` names are compatibility links containing the refreshed MiMo editions.
6
+
7
+ ## Reproduce
8
+
9
+ 1. Install Python 3.12 and Node.js 24. Run `python -m venv .venv`, then install `requirements.txt` into it.
10
+ 2. In `studio`, run `npm ci`. The project uses Remotion 4.0.536. Run `npm ci` in `transcriber` for local Whisper alignment.
11
+ 3. The included source audio and transcripts let you rebuild without a speech API. To synthesize again, set `MIMO_API_KEY` locally and run `python synthesize.py --all`. Existing audio is preserved, so move any segment you intend to regenerate to an archive first. MiMo generation is nondeterministic; reruns may differ.
12
+ 4. For new audio, run `node transcriber/transcribe.mjs` (configure its Python interpreter path if needed); remove only corresponding old transcript JSON files before rerunning. Existing transcript files are reused.
13
+ 5. Run `python engineering/assemble.py`, then `python engineering/register.py`.
14
+ 6. Run `python engineering/render.py films`. Set `WISP_CHROME` to your Chrome executable if needed. For interactive playback, run `npx remotion studio` inside `studio`.
15
+ 7. Run `python engineering/optimize_media.py`, then `python engineering/check_delivery.py` and review the alignment and media reports in `qa`.
16
+
17
+ Subtitles retain the source script, with timings aligned to local Whisper recognition. ASR spelling errors do not alter the script. The original English CIS subtitle splitting error is corrected. Minor pacing and loudness adjustments are recorded in `qa/alignment.json`.
18
+
19
+ The graphics and research claims were checked on 18 October 2026. This is a narration update, not a new experimental release.
20
+
21
+ ## Sources
22
+
23
+ - Speech API: https://mimo.mi.com/docs/en-US/quick-start/usage-guide/audio/speech-synthesis-v2.5
24
+ - Paper: https://research.birmingham.ac.uk/en/publications/wisp-composable-representations-for-wrist-worn-activity-recogniti/
25
+ - Method: https://huggingface.co/Zipeng365/WISP
26
+ - Benchmark: https://huggingface.co/datasets/Zipeng365/WristHARBench
videos/WISP_MiMo_reproducible_project.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:797c20c808bf1db672ca298d55c4029630d9ccb16f8f8cf78e99869052349a96
3
+ size 57091240
videos/WISP_en.srt CHANGED
@@ -1,235 +1,247 @@
1
  1
2
- 00:00:00,450 --> 00:00:02,424
3
- Walking, cooking, and handling objects
4
 
5
  2
6
- 00:00:02,424 --> 00:00:04,398
7
- leave different clues in wrist motion.
8
 
9
  3
10
- 00:00:04,638 --> 00:00:07,042
11
- Local shape, rhythm, temporal order, and
12
 
13
  4
14
- 00:00:07,042 --> 00:00:10,709
15
- relationships between axes can all help identify an activity.
16
 
17
  5
18
- 00:00:13,450 --> 00:00:15,595
19
- When benchmarks mix sensor locations, they
20
 
21
  6
22
- 00:00:15,595 --> 00:00:17,843
23
- cannot isolate what the wrist alone reveals.
24
 
25
  7
26
- 00:00:18,083 --> 00:00:20,178
27
- WristHARBench brings together nine wrist
28
 
29
  8
30
- 00:00:20,178 --> 00:00:22,902
31
- acceleration datasets under one evaluation contract.
32
 
33
  9
34
- 00:00:23,142 --> 00:00:25,031
35
- Two additional multichannel inertial
36
 
37
  10
38
- 00:00:25,031 --> 00:00:27,602
39
- datasets support a separate sensitivity analysis.
40
 
41
  11
42
- 00:00:30,450 --> 00:00:32,538
43
- Within each repeat, different people
44
 
45
  12
46
- 00:00:32,538 --> 00:00:35,091
47
- enter training, validation, and test groups.
48
 
49
  13
50
- 00:00:35,331 --> 00:00:37,231
51
- No person appears in two groups.
52
 
53
  14
54
- 00:00:37,231 --> 00:00:40,259
55
- Validation guides model choice; testing comes last.
56
 
57
  15
58
- 00:00:40,499 --> 00:00:44,406
59
- Methods share the same windows and splits, across five repeated holdouts.
60
 
61
  16
62
- 00:00:46,450 --> 00:00:50,120
63
- WISP reads the same signal window through complementary representations.
64
 
65
  17
66
- 00:00:50,360 --> 00:00:53,740
67
- Statistics and spectra describe overall variation and rhythm;
68
 
69
  18
70
- 00:00:53,740 --> 00:00:57,121
71
- other branches retain shape, order, and cross-axis structure.
72
 
73
  19
74
- 00:00:57,361 --> 00:01:00,377
75
- These landscapes illustrate operators on synthetic signals.
76
 
77
  20
78
- 00:01:01,683 --> 00:01:03,816
79
- Random convolution kernels move through
80
 
81
  21
82
- 00:01:03,816 --> 00:01:06,223
83
- time, responding to local waveform patterns.
84
 
85
  22
86
- 00:01:06,463 --> 00:01:09,757
87
- A selected slice reveals the responses within one signal window.
88
 
89
  23
90
- 00:01:09,997 --> 00:01:12,797
91
- Positive proportion, maximum, mean, and standard
92
 
93
  24
94
- 00:01:12,797 --> 00:01:15,597
95
- deviation pool that response into four features.
96
 
97
  25
98
- 00:01:19,683 --> 00:01:21,788
99
- The state-and-order representation maps
100
 
101
  26
102
- 00:01:21,788 --> 00:01:24,487
103
- values to relative, within-window quantile states.
104
 
105
  27
106
- 00:01:24,727 --> 00:01:27,870
107
- It counts how often states occur and how they transition to each other.
108
 
109
  28
110
- 00:01:28,110 --> 00:01:31,806
111
- The matrix contains transition probabilities, not attention weights.
112
 
113
  29
114
- 00:01:33,683 --> 00:01:36,487
115
- Geometry, interval, and shape features combine
116
 
117
  30
118
- 00:01:36,487 --> 00:01:40,389
119
- cross-axis structure, local statistics, and subsequence matches.
120
 
121
  31
122
- 00:01:40,629 --> 00:01:42,633
123
- A template moves along the window. The
124
 
125
  32
126
- 00:01:42,633 --> 00:01:44,847
127
- lowest distance identifies its best match.
128
 
129
  33
130
- 00:01:45,087 --> 00:01:47,587
131
- Both match quality and location become features.
132
 
133
  34
134
- 00:01:48,883 --> 00:01:51,624
135
- Complementary clues can be combined at the feature level.
136
 
137
  35
138
- 00:01:51,864 --> 00:01:55,855
139
- features I, and spectral, temporal, and cross-axis features S.
140
 
141
  36
142
- 00:01:55,855 --> 00:01:59,073
143
- features I, and spectral, temporal, and cross-axis
144
 
145
  37
146
- 00:01:59,313 --> 00:02:03,210
147
- A single classifier reads the combined vector and produces activity scores.
148
 
149
  38
150
- 00:02:04,883 --> 00:02:07,104
151
- Score ensembles combine independently fitted
152
 
153
  39
154
- 00:02:07,104 --> 00:02:09,425
155
- members after aligning their activity classes.
156
 
157
  40
158
- 00:02:09,665 --> 00:02:13,544
159
- CSE averages three members equally; ESE uses weighted averaging.
160
 
161
  41
162
- 00:02:13,784 --> 00:02:17,308
163
- This is different from concatenating features before one classifier.
164
 
165
  42
166
- 00:02:19,883 --> 00:02:22,641
167
- On the frozen paper evaluation, WISP with
168
 
169
  43
170
- 00:02:22,641 --> 00:02:26,004
171
- random search reaches 70.20 percent mean macro F1.
172
 
173
  44
174
- 00:02:26,244 --> 00:02:29,088
175
- RDST, the strongest of eighteen external
176
 
177
  45
178
- 00:02:29,088 --> 00:02:32,571
179
- baselines on this average, reaches 59.62 percent.
180
 
181
  46
182
- 00:02:32,811 --> 00:02:35,657
183
- The difference is 10.58 percentage points. This
184
 
185
  47
186
- 00:02:35,657 --> 00:02:38,623
187
- is a dataset average, not a win on every dataset.
188
 
189
  48
190
- 00:02:39,917 --> 00:02:43,940
191
- Controlled ablations test complementary information with a fixed classifier.
192
 
193
  49
194
- 00:02:44,180 --> 00:02:47,181
195
- With the other two groups already present,
196
 
197
  50
198
- 00:02:47,181 --> 00:02:50,610
199
- adding R gains 4.80 points; adding S gains 3.01.
200
 
201
  51
202
- 00:02:50,850 --> 00:02:53,367
203
- Adding I changes the mean by minus 0.27
204
 
205
  52
206
- 00:02:53,367 --> 00:02:55,883
207
- points, with an interval crossing zero.
208
 
209
  53
210
- 00:02:56,123 --> 00:02:58,322
211
- The evidence supports adding missing clues,
212
 
213
  54
214
- 00:02:58,322 --> 00:03:00,623
215
- rather than more overlapping representations.
216
 
217
  55
218
- 00:03:01,917 --> 00:03:04,910
219
- WristHARBench provides benchmark data. WISP provides
220
 
221
  56
222
- 00:03:04,910 --> 00:03:08,076
223
- method code, saved results, and retained search models.
224
 
225
  57
226
- 00:03:08,316 --> 00:03:10,712
227
- The 360 retained checkpoints do not cover
228
 
229
  58
230
- 00:03:10,712 --> 00:03:13,516
231
- every method or imply complete inference reruns.
232
 
233
  59
234
- 00:03:13,756 --> 00:03:15,325
 
 
 
 
 
 
 
 
 
 
 
 
235
  Explore the public releases.
 
1
  1
2
+ 00:00:01,004 --> 00:00:04,350
3
+ Walking, cooking, and handling objects leave
4
 
5
  2
6
+ 00:00:04,350 --> 00:00:06,715
7
+ different clues in wrist motion.
8
 
9
  3
10
+ 00:00:06,715 --> 00:00:12,042
11
+ Local shape, rhythm, temporal order, and relationships between axes can
12
 
13
  4
14
+ 00:00:12,042 --> 00:00:13,812
15
+ all help identify an activity.
16
 
17
  5
18
+ 00:00:15,168 --> 00:00:18,956
19
+ When benchmarks mix sensor locations, they cannot isolate
20
 
21
  6
22
+ 00:00:18,956 --> 00:00:21,022
23
+ what the wrist alone reveals.
24
 
25
  7
26
+ 00:00:21,022 --> 00:00:25,264
27
+ WristHARBench brings together nine wrist acceleration datasets
28
 
29
  8
30
+ 00:00:25,264 --> 00:00:27,841
31
+ under one evaluation contract.
32
 
33
  9
34
+ 00:00:27,841 --> 00:00:31,187
35
+ Two additional multichannel inertial datasets support a
36
 
37
  10
38
+ 00:00:31,187 --> 00:00:32,398
39
+ separate sensitivity analysis.
40
 
41
  11
42
+ 00:00:34,064 --> 00:00:37,718
43
+ Within each repeat, different people enter training,
44
 
45
  12
46
+ 00:00:37,718 --> 00:00:39,929
47
+ validation, and test groups.
48
 
49
  13
50
+ 00:00:39,929 --> 00:00:43,814
51
+ No person appears in two groups. Validation guides model
52
 
53
  14
54
+ 00:00:43,814 --> 00:00:46,679
55
+ choice; testing comes last.
56
 
57
  15
58
+ 00:00:46,679 --> 00:00:49,198
59
+ Methods share the same windows and splits,
60
 
61
  16
62
+ 00:00:49,198 --> 00:00:51,045
63
+ across five repeated holdouts.
64
 
65
  17
66
+ 00:00:52,296 --> 00:00:57,815
67
+ WISP reads the same signal window through complementary representations.
68
 
69
  18
70
+ 00:00:57,815 --> 00:01:02,412
71
+ Statistics and spectra describe overall variation and rhythm; other
72
 
73
  19
74
+ 00:01:02,412 --> 00:01:06,527
75
+ branches retain shape, order, and cross-axis structure.
76
 
77
  20
78
+ 00:01:06,527 --> 00:01:09,642
79
+ These landscapes illustrate operators on synthetic signals.
80
 
81
  21
82
+ 00:01:11,141 --> 00:01:14,487
83
+ Random convolution kernels move through time, responding
84
 
85
  22
86
+ 00:01:14,487 --> 00:01:16,679
87
+ to local waveform patterns.
88
 
89
  23
90
+ 00:01:16,679 --> 00:01:21,391
91
+ A selected slice reveals the responses within one signal window.
92
 
93
  24
94
+ 00:01:21,391 --> 00:01:26,737
95
+ Positive proportion, maximum, mean, and standard deviation pool that
96
 
97
  25
98
+ 00:01:26,737 --> 00:01:28,301
99
+ response into four features.
100
 
101
  26
102
+ 00:01:29,505 --> 00:01:33,775
103
+ The state-and-order representation maps values to relative,
104
 
105
  27
106
+ 00:01:33,775 --> 00:01:35,775
107
+ within-window quantile states.
108
 
109
  28
110
+ 00:01:35,775 --> 00:01:40,025
111
+ It counts how often states occur and how they transition to each other.
112
 
113
  29
114
+ 00:01:40,025 --> 00:01:44,026
115
+ The matrix contains transition probabilities, not attention weights.
116
 
117
  30
118
+ 00:01:45,527 --> 00:01:51,546
119
+ Geometry, interval, and shape features combine cross-axis structure,
120
 
121
  31
122
+ 00:01:51,546 --> 00:01:54,969
123
+ local statistics, and subsequence matches.
124
 
125
  32
126
+ 00:01:54,969 --> 00:01:58,854
127
+ A template moves along the window. The lowest distance
128
 
129
  33
130
+ 00:01:58,854 --> 00:02:00,835
131
+ identifies its best match.
132
 
133
  34
134
+ 00:02:00,835 --> 00:02:04,058
135
+ Both match quality and location become features.
136
 
137
  35
138
+ 00:02:05,271 --> 00:02:09,214
139
+ Complementary clues can be combined at the feature level.
140
 
141
  36
142
+ 00:02:09,214 --> 00:02:15,060
143
+ CIS concatenates convolution features R, interval features I, and
144
 
145
  37
146
+ 00:02:15,060 --> 00:02:18,771
147
+ spectral, temporal, and cross-axis features S.
148
 
149
  38
150
+ 00:02:18,771 --> 00:02:21,368
151
+ A single classifier reads the combined vector
152
 
153
  39
154
+ 00:02:21,368 --> 00:02:23,156
155
+ and produces activity scores.
156
 
157
  40
158
+ 00:02:24,333 --> 00:02:27,679
159
+ Score ensembles combine independently fitted members after
160
 
161
  41
162
+ 00:02:27,679 --> 00:02:29,775
163
+ aligning their activity classes.
164
 
165
  42
166
+ 00:02:29,775 --> 00:02:35,910
167
+ CSE averages three members equally; ESE uses weighted averaging.
168
 
169
  43
170
+ 00:02:35,910 --> 00:02:39,371
171
+ This is different from concatenating features before one classifier.
172
 
173
  44
174
+ 00:02:40,719 --> 00:02:45,508
175
+ On the frozen paper evaluation, WISP with random search reaches
176
 
177
  45
178
+ 00:02:45,508 --> 00:02:49,642
179
+ 70.20 percent mean macro F1.
180
 
181
  46
182
+ 00:02:49,642 --> 00:02:53,412
183
+ RDST, the strongest of eighteen external baselines on this
184
 
185
  47
186
+ 00:02:53,412 --> 00:02:57,027
187
+ average, reaches 59.62 percent.
188
 
189
  48
190
+ 00:02:57,027 --> 00:03:02,565
191
+ The difference is 10.58 percentage points. This is a dataset average,
192
 
193
  49
194
+ 00:03:02,565 --> 00:03:04,331
195
+ not a win on every dataset.
196
 
197
  50
198
+ 00:03:05,750 --> 00:03:07,962
199
+ Controlled ablations test complementary
200
 
201
  51
202
+ 00:03:07,962 --> 00:03:10,115
203
+ information with a fixed classifier.
204
 
205
  52
206
+ 00:03:10,115 --> 00:03:14,481
207
+ With the other two groups already present, adding R gains 4.80
208
 
209
  53
210
+ 00:03:14,481 --> 00:03:17,962
211
+ points; adding S gains 3.01.
212
 
213
  54
214
+ 00:03:17,962 --> 00:03:22,096
215
+ Adding I changes the mean by minus 0.27 points, with
216
 
217
  55
218
+ 00:03:22,096 --> 00:03:24,000
219
+ an interval crossing zero.
220
 
221
  56
222
+ 00:03:24,000 --> 00:03:27,308
223
+ The evidence supports adding missing clues, rather than more
224
 
225
  57
226
+ 00:03:27,308 --> 00:03:28,173
227
+ overlapping representations.
228
 
229
  58
230
+ 00:03:29,658 --> 00:03:34,523
231
+ WristHARBench provides benchmark data. WISP provides method code, saved
232
 
233
  59
234
+ 00:03:34,523 --> 00:03:36,965
235
+ results, and retained search models.
236
+
237
+ 60
238
+ 00:03:36,965 --> 00:03:42,369
239
+ The 360 retained checkpoints do not cover every method or imply
240
+
241
+ 61
242
+ 00:03:42,369 --> 00:03:44,312
243
+ complete inference reruns.
244
+
245
+ 62
246
+ 00:03:44,312 --> 00:03:45,496
247
  Explore the public releases.
videos/WISP_en.vtt CHANGED
@@ -1,178 +1,187 @@
1
  WEBVTT
2
 
3
- 00:00:00.450 --> 00:00:02.424
4
- Walking, cooking, and handling objects
5
 
6
- 00:00:02.424 --> 00:00:04.398
7
- leave different clues in wrist motion.
8
 
9
- 00:00:04.638 --> 00:00:07.042
10
- Local shape, rhythm, temporal order, and
11
 
12
- 00:00:07.042 --> 00:00:10.709
13
- relationships between axes can all help identify an activity.
14
 
15
- 00:00:13.450 --> 00:00:15.595
16
- When benchmarks mix sensor locations, they
17
 
18
- 00:00:15.595 --> 00:00:17.843
19
- cannot isolate what the wrist alone reveals.
20
 
21
- 00:00:18.083 --> 00:00:20.178
22
- WristHARBench brings together nine wrist
23
 
24
- 00:00:20.178 --> 00:00:22.902
25
- acceleration datasets under one evaluation contract.
26
 
27
- 00:00:23.142 --> 00:00:25.031
28
- Two additional multichannel inertial
29
 
30
- 00:00:25.031 --> 00:00:27.602
31
- datasets support a separate sensitivity analysis.
32
 
33
- 00:00:30.450 --> 00:00:32.538
34
- Within each repeat, different people
35
 
36
- 00:00:32.538 --> 00:00:35.091
37
- enter training, validation, and test groups.
38
 
39
- 00:00:35.331 --> 00:00:37.231
40
- No person appears in two groups.
41
 
42
- 00:00:37.231 --> 00:00:40.259
43
- Validation guides model choice; testing comes last.
44
 
45
- 00:00:40.499 --> 00:00:44.406
46
- Methods share the same windows and splits, across five repeated holdouts.
47
 
48
- 00:00:46.450 --> 00:00:50.120
 
 
 
49
  WISP reads the same signal window through complementary representations.
50
 
51
- 00:00:50.360 --> 00:00:53.740
52
- Statistics and spectra describe overall variation and rhythm;
53
 
54
- 00:00:53.740 --> 00:00:57.121
55
- other branches retain shape, order, and cross-axis structure.
56
 
57
- 00:00:57.361 --> 00:01:00.377
58
  These landscapes illustrate operators on synthetic signals.
59
 
60
- 00:01:01.683 --> 00:01:03.816
61
- Random convolution kernels move through
62
 
63
- 00:01:03.816 --> 00:01:06.223
64
- time, responding to local waveform patterns.
65
 
66
- 00:01:06.463 --> 00:01:09.757
67
  A selected slice reveals the responses within one signal window.
68
 
69
- 00:01:09.997 --> 00:01:12.797
70
- Positive proportion, maximum, mean, and standard
71
 
72
- 00:01:12.797 --> 00:01:15.597
73
- deviation pool that response into four features.
74
 
75
- 00:01:19.683 --> 00:01:21.788
76
- The state-and-order representation maps
77
 
78
- 00:01:21.788 --> 00:01:24.487
79
- values to relative, within-window quantile states.
80
 
81
- 00:01:24.727 --> 00:01:27.870
82
  It counts how often states occur and how they transition to each other.
83
 
84
- 00:01:28.110 --> 00:01:31.806
85
  The matrix contains transition probabilities, not attention weights.
86
 
87
- 00:01:33.683 --> 00:01:36.487
88
- Geometry, interval, and shape features combine
89
 
90
- 00:01:36.487 --> 00:01:40.389
91
- cross-axis structure, local statistics, and subsequence matches.
92
 
93
- 00:01:40.629 --> 00:01:42.633
94
- A template moves along the window. The
95
 
96
- 00:01:42.633 --> 00:01:44.847
97
- lowest distance identifies its best match.
98
 
99
- 00:01:45.087 --> 00:01:47.587
100
  Both match quality and location become features.
101
 
102
- 00:01:48.883 --> 00:01:51.624
103
  Complementary clues can be combined at the feature level.
104
 
105
- 00:01:51.864 --> 00:01:55.855
106
- features I, and spectral, temporal, and cross-axis features S.
 
 
 
107
 
108
- 00:01:55.855 --> 00:01:59.073
109
- features I, and spectral, temporal, and cross-axis
110
 
111
- 00:01:59.313 --> 00:02:03.210
112
- A single classifier reads the combined vector and produces activity scores.
113
 
114
- 00:02:04.883 --> 00:02:07.104
115
- Score ensembles combine independently fitted
116
 
117
- 00:02:07.104 --> 00:02:09.425
118
- members after aligning their activity classes.
119
 
120
- 00:02:09.665 --> 00:02:13.544
121
  CSE averages three members equally; ESE uses weighted averaging.
122
 
123
- 00:02:13.784 --> 00:02:17.308
124
  This is different from concatenating features before one classifier.
125
 
126
- 00:02:19.883 --> 00:02:22.641
127
- On the frozen paper evaluation, WISP with
 
 
 
128
 
129
- 00:02:22.641 --> 00:02:26.004
130
- random search reaches 70.20 percent mean macro F1.
131
 
132
- 00:02:26.244 --> 00:02:29.088
133
- RDST, the strongest of eighteen external
134
 
135
- 00:02:29.088 --> 00:02:32.571
136
- baselines on this average, reaches 59.62 percent.
137
 
138
- 00:02:32.811 --> 00:02:35.657
139
- The difference is 10.58 percentage points. This
140
 
141
- 00:02:35.657 --> 00:02:38.623
142
- is a dataset average, not a win on every dataset.
143
 
144
- 00:02:39.917 --> 00:02:43.940
145
- Controlled ablations test complementary information with a fixed classifier.
146
 
147
- 00:02:44.180 --> 00:02:47.181
148
- With the other two groups already present,
149
 
150
- 00:02:47.181 --> 00:02:50.610
151
- adding R gains 4.80 points; adding S gains 3.01.
152
 
153
- 00:02:50.850 --> 00:02:53.367
154
- Adding I changes the mean by minus 0.27
155
 
156
- 00:02:53.367 --> 00:02:55.883
157
- points, with an interval crossing zero.
158
 
159
- 00:02:56.123 --> 00:02:58.322
160
- The evidence supports adding missing clues,
161
 
162
- 00:02:58.322 --> 00:03:00.623
163
- rather than more overlapping representations.
164
 
165
- 00:03:01.917 --> 00:03:04.910
166
- WristHARBench provides benchmark data. WISP provides
167
 
168
- 00:03:04.910 --> 00:03:08.076
169
- method code, saved results, and retained search models.
170
 
171
- 00:03:08.316 --> 00:03:10.712
172
- The 360 retained checkpoints do not cover
173
 
174
- 00:03:10.712 --> 00:03:13.516
175
- every method or imply complete inference reruns.
176
 
177
- 00:03:13.756 --> 00:03:15.325
178
  Explore the public releases.
 
1
  WEBVTT
2
 
3
+ 00:00:01.004 --> 00:00:04.350
4
+ Walking, cooking, and handling objects leave
5
 
6
+ 00:00:04.350 --> 00:00:06.715
7
+ different clues in wrist motion.
8
 
9
+ 00:00:06.715 --> 00:00:12.042
10
+ Local shape, rhythm, temporal order, and relationships between axes can
11
 
12
+ 00:00:12.042 --> 00:00:13.812
13
+ all help identify an activity.
14
 
15
+ 00:00:15.168 --> 00:00:18.956
16
+ When benchmarks mix sensor locations, they cannot isolate
17
 
18
+ 00:00:18.956 --> 00:00:21.022
19
+ what the wrist alone reveals.
20
 
21
+ 00:00:21.022 --> 00:00:25.264
22
+ WristHARBench brings together nine wrist acceleration datasets
23
 
24
+ 00:00:25.264 --> 00:00:27.841
25
+ under one evaluation contract.
26
 
27
+ 00:00:27.841 --> 00:00:31.187
28
+ Two additional multichannel inertial datasets support a
29
 
30
+ 00:00:31.187 --> 00:00:32.398
31
+ separate sensitivity analysis.
32
 
33
+ 00:00:34.064 --> 00:00:37.718
34
+ Within each repeat, different people enter training,
35
 
36
+ 00:00:37.718 --> 00:00:39.929
37
+ validation, and test groups.
38
 
39
+ 00:00:39.929 --> 00:00:43.814
40
+ No person appears in two groups. Validation guides model
41
 
42
+ 00:00:43.814 --> 00:00:46.679
43
+ choice; testing comes last.
44
 
45
+ 00:00:46.679 --> 00:00:49.198
46
+ Methods share the same windows and splits,
47
 
48
+ 00:00:49.198 --> 00:00:51.045
49
+ across five repeated holdouts.
50
+
51
+ 00:00:52.296 --> 00:00:57.815
52
  WISP reads the same signal window through complementary representations.
53
 
54
+ 00:00:57.815 --> 00:01:02.412
55
+ Statistics and spectra describe overall variation and rhythm; other
56
 
57
+ 00:01:02.412 --> 00:01:06.527
58
+ branches retain shape, order, and cross-axis structure.
59
 
60
+ 00:01:06.527 --> 00:01:09.642
61
  These landscapes illustrate operators on synthetic signals.
62
 
63
+ 00:01:11.141 --> 00:01:14.487
64
+ Random convolution kernels move through time, responding
65
 
66
+ 00:01:14.487 --> 00:01:16.679
67
+ to local waveform patterns.
68
 
69
+ 00:01:16.679 --> 00:01:21.391
70
  A selected slice reveals the responses within one signal window.
71
 
72
+ 00:01:21.391 --> 00:01:26.737
73
+ Positive proportion, maximum, mean, and standard deviation pool that
74
 
75
+ 00:01:26.737 --> 00:01:28.301
76
+ response into four features.
77
 
78
+ 00:01:29.505 --> 00:01:33.775
79
+ The state-and-order representation maps values to relative,
80
 
81
+ 00:01:33.775 --> 00:01:35.775
82
+ within-window quantile states.
83
 
84
+ 00:01:35.775 --> 00:01:40.025
85
  It counts how often states occur and how they transition to each other.
86
 
87
+ 00:01:40.025 --> 00:01:44.026
88
  The matrix contains transition probabilities, not attention weights.
89
 
90
+ 00:01:45.527 --> 00:01:51.546
91
+ Geometry, interval, and shape features combine cross-axis structure,
92
 
93
+ 00:01:51.546 --> 00:01:54.969
94
+ local statistics, and subsequence matches.
95
 
96
+ 00:01:54.969 --> 00:01:58.854
97
+ A template moves along the window. The lowest distance
98
 
99
+ 00:01:58.854 --> 00:02:00.835
100
+ identifies its best match.
101
 
102
+ 00:02:00.835 --> 00:02:04.058
103
  Both match quality and location become features.
104
 
105
+ 00:02:05.271 --> 00:02:09.214
106
  Complementary clues can be combined at the feature level.
107
 
108
+ 00:02:09.214 --> 00:02:15.060
109
+ CIS concatenates convolution features R, interval features I, and
110
+
111
+ 00:02:15.060 --> 00:02:18.771
112
+ spectral, temporal, and cross-axis features S.
113
 
114
+ 00:02:18.771 --> 00:02:21.368
115
+ A single classifier reads the combined vector
116
 
117
+ 00:02:21.368 --> 00:02:23.156
118
+ and produces activity scores.
119
 
120
+ 00:02:24.333 --> 00:02:27.679
121
+ Score ensembles combine independently fitted members after
122
 
123
+ 00:02:27.679 --> 00:02:29.775
124
+ aligning their activity classes.
125
 
126
+ 00:02:29.775 --> 00:02:35.910
127
  CSE averages three members equally; ESE uses weighted averaging.
128
 
129
+ 00:02:35.910 --> 00:02:39.371
130
  This is different from concatenating features before one classifier.
131
 
132
+ 00:02:40.719 --> 00:02:45.508
133
+ On the frozen paper evaluation, WISP with random search reaches
134
+
135
+ 00:02:45.508 --> 00:02:49.642
136
+ 70.20 percent mean macro F1.
137
 
138
+ 00:02:49.642 --> 00:02:53.412
139
+ RDST, the strongest of eighteen external baselines on this
140
 
141
+ 00:02:53.412 --> 00:02:57.027
142
+ average, reaches 59.62 percent.
143
 
144
+ 00:02:57.027 --> 00:03:02.565
145
+ The difference is 10.58 percentage points. This is a dataset average,
146
 
147
+ 00:03:02.565 --> 00:03:04.331
148
+ not a win on every dataset.
149
 
150
+ 00:03:05.750 --> 00:03:07.962
151
+ Controlled ablations test complementary
152
 
153
+ 00:03:07.962 --> 00:03:10.115
154
+ information with a fixed classifier.
155
 
156
+ 00:03:10.115 --> 00:03:14.481
157
+ With the other two groups already present, adding R gains 4.80
158
 
159
+ 00:03:14.481 --> 00:03:17.962
160
+ points; adding S gains 3.01.
161
 
162
+ 00:03:17.962 --> 00:03:22.096
163
+ Adding I changes the mean by minus 0.27 points, with
164
 
165
+ 00:03:22.096 --> 00:03:24.000
166
+ an interval crossing zero.
167
 
168
+ 00:03:24.000 --> 00:03:27.308
169
+ The evidence supports adding missing clues, rather than more
170
 
171
+ 00:03:27.308 --> 00:03:28.173
172
+ overlapping representations.
173
 
174
+ 00:03:29.658 --> 00:03:34.523
175
+ WristHARBench provides benchmark data. WISP provides method code, saved
176
 
177
+ 00:03:34.523 --> 00:03:36.965
178
+ results, and retained search models.
179
 
180
+ 00:03:36.965 --> 00:03:42.369
181
+ The 360 retained checkpoints do not cover every method or imply
182
 
183
+ 00:03:42.369 --> 00:03:44.312
184
+ complete inference reruns.
185
 
186
+ 00:03:44.312 --> 00:03:45.496
187
  Explore the public releases.
videos/WISP_en_mimo_voice.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6a84479dd6c90c65300f8a681c7dcf3a9cc2ef4cb023512f3ebf3924f3596cb1
3
+ size 102486414
videos/WISP_en_mimo_voice.wav ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5707f14bbc83074c5aeea409c88faa1faed9ad6d7c804cd8857d7150c1e80081
3
+ size 21702444
videos/WISP_en_script.txt CHANGED
@@ -1,82 +1,59 @@
1
  00:00:00,000 Reading activity at the wrist
2
- Walking, cooking, and handling objects
3
- leave different clues in wrist motion.
4
- Local shape, rhythm, temporal order, and
5
- relationships between axes can all help identify an activity.
6
 
7
- 00:00:13,000 WristHARBench: a common comparison
8
- When benchmarks mix sensor locations, they
9
- cannot isolate what the wrist alone reveals.
10
- WristHARBench brings together nine wrist
11
- acceleration datasets under one evaluation contract.
12
- Two additional multichannel inertial
13
- datasets support a separate sensitivity analysis.
14
 
15
- 00:00:30,000 Separate people before comparing models
16
- Within each repeat, different people
17
- enter training, validation, and test groups.
18
- No person appears in two groups.
19
- Validation guides model choice; testing comes last.
20
  Methods share the same windows and splits, across five repeated holdouts.
21
 
22
- 00:00:46,000 One window, several readings
23
  WISP reads the same signal window through complementary representations.
24
- Statistics and spectra describe overall variation and rhythm;
25
- other branches retain shape, order, and cross-axis structure.
26
  These landscapes illustrate operators on synthetic signals.
27
 
28
- 00:01:01,233 RC: from responses to features
29
- Random convolution kernels move through
30
- time, responding to local waveform patterns.
31
  A selected slice reveals the responses within one signal window.
32
- Positive proportion, maximum, mean, and standard
33
- deviation pool that response into four features.
34
 
35
- 00:01:19,233 SO: relative states and their order
36
- The state-and-order representation maps
37
- values to relative, within-window quantile states.
38
  It counts how often states occur and how they transition to each other.
39
  The matrix contains transition probabilities, not attention weights.
40
 
41
- 00:01:33,233 GIS: finding shapes within a window
42
- Geometry, interval, and shape features combine
43
- cross-axis structure, local statistics, and subsequence matches.
44
- A template moves along the window. The
45
- lowest distance identifies its best match.
46
  Both match quality and location become features.
47
 
48
- 00:01:48,433 CIS: concatenate features first
49
  Complementary clues can be combined at the feature level.
50
- features I, and spectral, temporal, and cross-axis features S.
51
- features I, and spectral, temporal, and cross-axis
52
  A single classifier reads the combined vector and produces activity scores.
53
 
54
- 00:02:04,433 CSE / ESE: combine class scores
55
- Score ensembles combine independently fitted
56
- members after aligning their activity classes.
57
  CSE averages three members equally; ESE uses weighted averaging.
58
  This is different from concatenating features before one classifier.
59
 
60
- 00:02:19,433 Nine datasets, equal weight
61
- On the frozen paper evaluation, WISP with
62
- random search reaches 70.20 percent mean macro F1.
63
- RDST, the strongest of eighteen external
64
- baselines on this average, reaches 59.62 percent.
65
- The difference is 10.58 percentage points. This
66
- is a dataset average, not a win on every dataset.
67
 
68
- 00:02:39,467 What information does a branch add?
69
  Controlled ablations test complementary information with a fixed classifier.
70
- With the other two groups already present,
71
- adding R gains 4.80 points; adding S gains 3.01.
72
- Adding I changes the mean by minus 0.27
73
- points, with an interval crossing zero.
74
- The evidence supports adding missing clues,
75
- rather than more overlapping representations.
76
-
77
- 00:03:01,467 From shared evaluation to composable representations
78
- WristHARBench provides benchmark data. WISP provides
79
- method code, saved results, and retained search models.
80
- The 360 retained checkpoints do not cover
81
- every method or imply complete inference reruns.
82
- Explore the public releases.
 
1
  00:00:00,000 Reading activity at the wrist
2
+ Walking, cooking, and handling objects leave different clues in wrist motion.
3
+ Local shape, rhythm, temporal order, and relationships between axes can all help identify an activity.
 
 
4
 
5
+ 00:00:14,433 WristHARBench: a common comparison
6
+ When benchmarks mix sensor locations, they cannot isolate what the wrist alone reveals.
7
+ WristHARBench brings together nine wrist acceleration datasets under one evaluation contract.
8
+ Two additional multichannel inertial datasets support a separate sensitivity analysis.
 
 
 
9
 
10
+ 00:00:33,233 Separate people before comparing models
11
+ Within each repeat, different people enter training, validation, and test groups.
12
+ No person appears in two groups. Validation guides model choice; testing comes last.
 
 
13
  Methods share the same windows and splits, across five repeated holdouts.
14
 
15
+ 00:00:51,600 One window, several readings
16
  WISP reads the same signal window through complementary representations.
17
+ Statistics and spectra describe overall variation and rhythm; other branches retain shape, order, and cross-axis structure.
 
18
  These landscapes illustrate operators on synthetic signals.
19
 
20
+ 00:01:10,233 RC: from responses to features
21
+ Random convolution kernels move through time, responding to local waveform patterns.
 
22
  A selected slice reveals the responses within one signal window.
23
+ Positive proportion, maximum, mean, and standard deviation pool that response into four features.
 
24
 
25
+ 00:01:28,867 SO: relative states and their order
26
+ The state-and-order representation maps values to relative, within-window quantile states.
 
27
  It counts how often states occur and how they transition to each other.
28
  The matrix contains transition probabilities, not attention weights.
29
 
30
+ 00:01:44,600 GIS: finding shapes within a window
31
+ Geometry, interval, and shape features combine cross-axis structure, local statistics, and subsequence matches.
32
+ A template moves along the window. The lowest distance identifies its best match.
 
 
33
  Both match quality and location become features.
34
 
35
+ 00:02:04,633 CIS: concatenate features first
36
  Complementary clues can be combined at the feature level.
37
+ CIS concatenates convolution features R, interval features I, and spectral, temporal, and cross-axis features S.
 
38
  A single classifier reads the combined vector and produces activity scores.
39
 
40
+ 00:02:23,733 CSE / ESE: combine class scores
41
+ Score ensembles combine independently fitted members after aligning their activity classes.
 
42
  CSE averages three members equally; ESE uses weighted averaging.
43
  This is different from concatenating features before one classifier.
44
 
45
+ 00:02:40,100 Nine datasets, equal weight
46
+ On the frozen paper evaluation, WISP with random search reaches 70.20 percent mean macro F1.
47
+ RDST, the strongest of eighteen external baselines on this average, reaches 59.62 percent.
48
+ The difference is 10.58 percentage points. This is a dataset average, not a win on every dataset.
 
 
 
49
 
50
+ 00:03:04,900 What information does a branch add?
51
  Controlled ablations test complementary information with a fixed classifier.
52
+ With the other two groups already present, adding R gains 4.80 points; adding S gains 3.01.
53
+ Adding I changes the mean by minus 0.27 points, with an interval crossing zero.
54
+ The evidence supports adding missing clues, rather than more overlapping representations.
55
+
56
+ 00:03:29,000 From shared evaluation to composable representations
57
+ WristHARBench provides benchmark data. WISP provides method code, saved results, and retained search models.
58
+ The 360 retained checkpoints do not cover every method or imply complete inference reruns.
59
+ Explore the public releases.
 
 
 
 
 
videos/WISP_en_system_voice.mp4 CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b1d36676464b1eecd714802f6be25c770074b682f510a2142db2b55b7e4822c8
3
- size 90249749
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6a84479dd6c90c65300f8a681c7dcf3a9cc2ef4cb023512f3ebf3924f3596cb1
3
+ size 102486414
videos/WISP_zh.srt CHANGED
@@ -1,183 +1,187 @@
1
  1
2
- 00:00:00,450 --> 00:00:05,905
3
  同样是手腕运动,走路、做饭和整理物品,留下的线索并不相同。
4
 
5
  2
6
- 00:00:06,145 --> 00:00:10,439
7
  局部形状、节律、先后次序,还有不同轴之间的关系,
8
 
9
  3
10
- 00:00:10,439 --> 00:00:12,229
11
  都可能帮助识别活动。
12
 
13
  4
14
- 00:00:13,550 --> 00:00:19,077
15
- 如果把身体不同部位的传感器混在一起,结果就难以回答:只看手腕,
16
 
17
  5
18
- 00:00:19,077 --> 00:00:20,503
19
- 究竟能识别什么?
20
 
21
  6
22
- 00:00:20,743 --> 00:00:24,161
23
  WristHARBench 整理了九个腕部三轴加速度数据集。
24
 
25
  7
26
- 00:00:24,401 --> 00:00:27,664
27
  另有两个多通道惯性数据集,单独分析。
28
 
29
  8
30
- 00:00:30,550 --> 00:00:34,744
31
  每次划分中,不同参与者进入训练、验证和测试组,
32
 
33
  9
34
- 00:00:34,744 --> 00:00:36,750
35
  同一个人不会跨组出现。
36
 
37
  10
38
- 00:00:36,990 --> 00:00:40,532
39
  模型设置只用验证组选择,测试组留到最后。
40
 
41
  11
42
- 00:00:40,772 --> 00:00:44,503
43
  所有方法使用相同的窗口和划分,重复五次。
44
 
45
  12
46
- 00:00:46,550 --> 00:00:49,420
47
  WISP 把同一段信号转换为多种表征。
48
 
49
  13
50
- 00:00:49,660 --> 00:00:52,435
51
  统计和频谱描述整体变化与节律;
52
 
53
  14
54
- 00:00:52,435 --> 00:00:56,319
55
  其他分支保留局部形状、相对次序和轴间关系。
56
 
57
  15
58
- 00:00:56,559 --> 00:00:59,751
59
  这里的曲面来自合成信号,是机制示意。
60
 
61
  16
62
- 00:01:01,050 --> 00:01:05,351
63
  随机卷积核沿时间移动,对不同的局部波形产生响应。
64
 
65
  17
66
- 00:01:05,591 --> 00:01:09,659
67
  选定一条截线,就能看到同一窗口内的响应变化。
68
 
69
  18
70
- 00:01:09,899 --> 00:01:15,345
71
  正值比例、最大值、均值和标准差,把整条响应汇总为四个特征。
72
 
73
  19
74
- 00:01:19,050 --> 00:01:23,868
75
  状态与次序表征,把窗口内的数值按分位数转成相对状态。
76
 
77
  20
78
- 00:01:24,108 --> 00:01:28,133
79
  再统计各状态出现多少,以及它们如何相互转移。
80
 
81
  21
82
- 00:01:28,373 --> 00:01:31,379
83
  矩阵记录的是转移概率,不是注意力权重。
84
 
85
  22
86
- 00:01:33,050 --> 00:01:38,900
87
- 几何、区间和形状表征,同时读取轴间结构、局部统计和子序列匹配。
88
 
89
  23
90
- 00:01:39,140 --> 00:01:42,769
91
- 模板沿窗口移动,距离最低的位置给出最佳匹配。
92
 
93
  24
94
- 00:01:43,009 --> 00:01:46,571
95
- 保留下来的线索,包括匹配质量和出现位置。
96
 
97
  25
98
- 00:01:47,883 --> 00:01:50,102
99
- 互补线索可以在特征层组合。
100
 
101
  26
102
- 00:01:50,342 --> 00:01:53,625
103
- CIS 把卷积特征 R、区间特征 I,
104
 
105
  27
106
- 00:01:53,625 --> 00:01:57,600
107
- 以及频谱、时间和轴间特征 S,拼接成一个向量。
108
 
109
  28
110
- 00:01:57,840 --> 00:02:01,451
111
- 随后由同一个分类器读取,输出各活动的分数。
112
 
113
  29
114
- 00:02:03,883 --> 00:02:07,747
115
- 另一种组合发生在分数层:各成员先独立预测,
116
 
117
  30
118
- 00:02:07,747 --> 00:02:09,955
119
- 再按相同的活动顺序对齐。
120
 
121
  31
122
- 00:02:10,195 --> 00:02:14,525
123
- CSE 等权平均三个成员,ESE 使用加权平均。
124
 
125
  32
126
- 00:02:14,765 --> 00:02:17,912
127
- 这与特征拼接,是两种不同的组合方式。
128
 
129
  33
130
- 00:02:19,217 --> 00:02:20,903
131
- 在冻结的论文评估范围内,
132
 
133
  34
134
- 00:02:20,903 --> 00:02:25,682
135
- 随机搜索版 WISP 的平均 Macro F1 为百分之七十点二零。
136
 
137
  35
138
- 00:02:25,922 --> 00:02:30,753
139
- 十八个外部基线中,最高的 RDST 为百分之五十九点六二,
140
 
141
  36
142
- 00:02:30,753 --> 00:02:32,586
143
- 相差十点五八个百分点。
144
 
145
  37
146
- 00:02:32,826 --> 00:02:37,090
147
- 这是九个数据集的平均结果,不代表每个数据集都获胜。
148
 
149
  38
150
- 00:02:38,383 --> 00:02:40,657
151
- 受控消融进一步检验互补性。
152
 
153
  39
154
- 00:02:40,897 --> 00:02:44,778
155
- 当另外两组特征已经存在时,加入 R 和 S,
156
 
157
  40
158
- 00:02:44,778 --> 00:02:48,307
159
- 分别平均增加四点八零和三点零一个百分点。
160
 
161
  41
162
- 00:02:48,547 --> 00:02:52,390
163
- 加入 I 的平均变化为负零点二七,区间跨过零。
164
 
165
  42
166
- 00:02:52,630 --> 00:02:56,893
167
- 结果支持补充缺失线索,而不是不断增加重复表征。
168
 
169
  43
170
- 00:02:58,217 --> 00:03:01,034
171
- WristHARBench 提供基准数据;
172
 
173
  44
174
- 00:03:01,034 --> 00:03:04,656
175
- WISP 提供方法代码、保存结果和实际保留的搜索模型。
176
 
177
  45
178
- 00:03:04,896 --> 00:03:09,297
179
- 三百六十个检查点不等于全部模型,也不代表完整复测。
180
 
181
  46
182
- 00:03:09,537 --> 00:03:11,598
 
 
 
 
183
  从公开入口,继续探索。
 
1
  1
2
+ 00:00:00,650 --> 00:00:07,868
3
  同样是手腕运动,走路、做饭和整理物品,留下的线索并不相同。
4
 
5
  2
6
+ 00:00:07,868 --> 00:00:12,396
7
  局部形状、节律、先后次序,还有不同轴之间的关系,
8
 
9
  3
10
+ 00:00:12,396 --> 00:00:13,796
11
  都可能帮助识别活动。
12
 
13
  4
14
+ 00:00:15,015 --> 00:00:18,725
15
+ 如果把身体不同部位的传感器混在一起,
16
 
17
  5
18
+ 00:00:18,725 --> 00:00:23,541
19
+ 结果就难以回答:只看手腕,究竟能识别什么?
20
 
21
  6
22
+ 00:00:23,541 --> 00:00:28,543
23
  WristHARBench 整理了九个腕部三轴加速度数据集。
24
 
25
  7
26
+ 00:00:28,543 --> 00:00:31,560
27
  另有两个多通道惯性数据集,单独分析。
28
 
29
  8
30
+ 00:00:32,721 --> 00:00:38,231
31
  每次划分中,不同参与者进入训练、验证和测试组,
32
 
33
  9
34
+ 00:00:38,231 --> 00:00:40,885
35
  同一个人不会跨组出现。
36
 
37
  10
38
+ 00:00:40,885 --> 00:00:45,776
39
  模型设置只用验证组选择,测试组留到最后。
40
 
41
  11
42
+ 00:00:45,776 --> 00:00:49,406
43
  所有方法使用相同的窗口和划分,重复五次。
44
 
45
  12
46
+ 00:00:50,709 --> 00:00:56,600
47
  WISP 把同一段信号转换为多种表征。
48
 
49
  13
50
+ 00:00:56,600 --> 00:01:00,837
51
  统计和频谱描述整体变化与节律;
52
 
53
  14
54
+ 00:01:00,837 --> 00:01:06,764
55
  其他分支保留局部形状、相对次序和轴间关系。
56
 
57
  15
58
+ 00:01:06,764 --> 00:01:10,657
59
  这里的曲面来自合成信号,是机制示意。
60
 
61
  16
62
+ 00:01:11,865 --> 00:01:18,138
63
  随机卷积核沿时间移动,对不同的局部波形产生响应。
64
 
65
  17
66
+ 00:01:18,138 --> 00:01:23,120
67
  选定一条截线,就能看到同一窗口内的响应变化。
68
 
69
  18
70
+ 00:01:23,120 --> 00:01:29,938
71
  正值比例、最大值、均值和标准差,把整条响应汇总为四个特征。
72
 
73
  19
74
+ 00:01:31,146 --> 00:01:38,964
75
  状态与次序表征,把窗口内的数值按分位数转成相对状态。
76
 
77
  20
78
+ 00:01:38,964 --> 00:01:44,046
79
  再统计各状态出现多少,以及它们如何相互转移。
80
 
81
  21
82
+ 00:01:44,046 --> 00:01:48,917
83
  矩阵记录的是转移概率,不是注意力权重。
84
 
85
  22
86
+ 00:01:50,104 --> 00:01:53,231
87
+ 几何、区间和形状表征,
88
 
89
  23
90
+ 00:01:53,231 --> 00:01:58,849
91
+ 同时读取轴间结构、局部统计和子序列匹配。
92
 
93
  24
94
+ 00:01:58,849 --> 00:02:03,940
95
+ 模板沿窗口移动,距离最低的位置给出最佳匹配。
96
 
97
  25
98
+ 00:02:03,940 --> 00:02:07,558
99
+ 保留下来的线索,包括匹配质量和出现位置。
100
 
101
  26
102
+ 00:02:08,735 --> 00:02:12,590
103
+ 互补线索可以在特征层组合。
104
 
105
  27
106
+ 00:02:12,590 --> 00:02:16,844
107
+ CIS 把卷积特征 R、区间特征 I,
108
 
109
  28
110
+ 00:02:16,844 --> 00:02:21,826
111
+ 以及频谱、时间和轴间特征 S,拼接成一个向量。
112
 
113
  29
114
+ 00:02:21,826 --> 00:02:25,572
115
+ 随后由同一个分类器读取,输出各活动的分数。
116
 
117
  30
118
+ 00:02:26,918 --> 00:02:31,791
119
+ 另一种组合发生在分数层:各成员先独立预测,
120
 
121
  31
122
+ 00:02:31,791 --> 00:02:34,882
123
+ 再按相同的活动顺序对齐。
124
 
125
  32
126
+ 00:02:34,882 --> 00:02:40,046
127
+ CSE 等权平均三个成员,ESE 使用加权平均。
128
 
129
  33
130
+ 00:02:40,046 --> 00:02:42,936
131
+ 这与特征拼接,是两种不同的组合方式。
132
 
133
  34
134
+ 00:02:44,409 --> 00:02:46,864
135
+ 在冻结的论文评估范围内,
136
 
137
  35
138
+ 00:02:46,864 --> 00:02:53,421
139
+ 随机搜索版 WISP 的平均 Macro F1 为百分之七十点二零。
140
 
141
  36
142
+ 00:02:53,421 --> 00:02:58,573
143
+ 十八个外部基线中,最高的 RDST 为百分之五十九点六二,
144
 
145
  37
146
+ 00:02:58,573 --> 00:03:01,955
147
+ 相差十点五八个百分点。
148
 
149
  38
150
+ 00:03:01,955 --> 00:03:06,155
151
+ 这是九个数据集的平均结果,不代表每个数据集都获胜。
152
 
153
  39
154
+ 00:03:07,419 --> 00:03:11,110
155
+ 受控消融进一步检验互补性。
156
 
157
  40
158
+ 00:03:11,110 --> 00:03:15,256
159
+ 当另外两组特征已经存在时,加入 R 和 S,
160
 
161
  41
162
+ 00:03:15,256 --> 00:03:20,419
163
+ 分别平均增加四点八零和三点零一个百分点。
164
 
165
  42
166
+ 00:03:20,419 --> 00:03:25,801
167
+ 加入 I 的平均变化为负零点二七,区间跨过零。
168
 
169
  43
170
+ 00:03:25,801 --> 00:03:30,365
171
+ 结果支持补充缺失线索,而不是不断增加重复表征。
172
 
173
  44
174
+ 00:03:31,505 --> 00:03:34,396
175
+ WristHARBench 提供基准数据;
176
 
177
  45
178
+ 00:03:34,396 --> 00:03:39,719
179
+ WISP 提供方法代码、保存结果和实际保留的搜索模型。
180
 
181
  46
182
+ 00:03:39,719 --> 00:03:44,778
183
+ 三百六十个检查点不等于全部模型,也不代表完整复测。
184
+
185
+ 47
186
+ 00:03:44,778 --> 00:03:46,542
187
  从公开入口,继续探索。
videos/WISP_zh.vtt CHANGED
@@ -1,139 +1,142 @@
1
  WEBVTT
2
 
3
- 00:00:00.450 --> 00:00:05.905
4
  同样是手腕运动,走路、做饭和整理物品,留下的线索并不相同。
5
 
6
- 00:00:06.145 --> 00:00:10.439
7
  局部形状、节律、先后次序,还有不同轴之间的关系,
8
 
9
- 00:00:10.439 --> 00:00:12.229
10
  都可能帮助识别活动。
11
 
12
- 00:00:13.550 --> 00:00:19.077
13
- 如果把身体不同部位的传感器混在一起,结果就难以回答:只看手腕,
14
 
15
- 00:00:19.077 --> 00:00:20.503
16
- 究竟能识别什么?
17
 
18
- 00:00:20.743 --> 00:00:24.161
19
  WristHARBench 整理了九个腕部三轴加速度数据集。
20
 
21
- 00:00:24.401 --> 00:00:27.664
22
  另有两个多通道惯性数据集,单独分析。
23
 
24
- 00:00:30.550 --> 00:00:34.744
25
  每次划分中,不同参与者进入训练、验证和测试组,
26
 
27
- 00:00:34.744 --> 00:00:36.750
28
  同一个人不会跨组出现。
29
 
30
- 00:00:36.990 --> 00:00:40.532
31
  模型设置只用验证组选择,测试组留到最后。
32
 
33
- 00:00:40.772 --> 00:00:44.503
34
  所有方法使用相同的窗口和划分,重复五次。
35
 
36
- 00:00:46.550 --> 00:00:49.420
37
  WISP 把同一段信号转换为多种表征。
38
 
39
- 00:00:49.660 --> 00:00:52.435
40
  统计和频谱描述整体变化与节律;
41
 
42
- 00:00:52.435 --> 00:00:56.319
43
  其他分支保留局部形状、相对次序和轴间关系。
44
 
45
- 00:00:56.559 --> 00:00:59.751
46
  这里的曲面来自合成信号,是机制示意。
47
 
48
- 00:01:01.050 --> 00:01:05.351
49
  随机卷积核沿时间移动,对不同的局部波形产生响应。
50
 
51
- 00:01:05.591 --> 00:01:09.659
52
  选定一条截线,就能看到同一窗口内的响应变化。
53
 
54
- 00:01:09.899 --> 00:01:15.345
55
  正值比例、最大值、均值和标准差,把整条响应汇总为四个特征。
56
 
57
- 00:01:19.050 --> 00:01:23.868
58
  状态与次序表征,把窗口内的数值按分位数转成相对状态。
59
 
60
- 00:01:24.108 --> 00:01:28.133
61
  再统计各状态出现多少,以及它们如何相互转移。
62
 
63
- 00:01:28.373 --> 00:01:31.379
64
  矩阵记录的是转移概率,不是注意力权重。
65
 
66
- 00:01:33.050 --> 00:01:38.900
67
- 几何、区间和形状表征,同时读取轴间结构、局部统计和子序列匹配。
68
 
69
- 00:01:39.140 --> 00:01:42.769
 
 
 
70
  模板沿窗口移动,距离最低的位置给出最佳匹配。
71
 
72
- 00:01:43.009 --> 00:01:46.571
73
  保留下来的线索,包括匹配质量和出现位置。
74
 
75
- 00:01:47.883 --> 00:01:50.102
76
  互补线索可以在特征层组合。
77
 
78
- 00:01:50.342 --> 00:01:53.625
79
  CIS 把卷积特征 R、区间特征 I,
80
 
81
- 00:01:53.625 --> 00:01:57.600
82
  以及频谱、时间和轴间特征 S,拼接成一个向量。
83
 
84
- 00:01:57.840 --> 00:02:01.451
85
  随后由同一个分类器读取,输出各活动的分数。
86
 
87
- 00:02:03.883 --> 00:02:07.747
88
  另一种组合发生在分数层:各成员先独立预测,
89
 
90
- 00:02:07.747 --> 00:02:09.955
91
  再按相同的活动顺序对齐。
92
 
93
- 00:02:10.195 --> 00:02:14.525
94
  CSE 等权平均三个成员,ESE 使用加权平均。
95
 
96
- 00:02:14.765 --> 00:02:17.912
97
  这与特征拼接,是两种不同的组合方式。
98
 
99
- 00:02:19.217 --> 00:02:20.903
100
  在冻结的论文评估范围内,
101
 
102
- 00:02:20.903 --> 00:02:25.682
103
  随机搜索版 WISP 的平均 Macro F1 为百分之七十点二零。
104
 
105
- 00:02:25.922 --> 00:02:30.753
106
  十八个外部基线中,最高的 RDST 为百分之五十九点六二,
107
 
108
- 00:02:30.753 --> 00:02:32.586
109
  相差十点五八个百分点。
110
 
111
- 00:02:32.826 --> 00:02:37.090
112
  这是九个数据集的平均结果,不代表每个数据集都获胜。
113
 
114
- 00:02:38.383 --> 00:02:40.657
115
  受控消融进一步检验互补性。
116
 
117
- 00:02:40.897 --> 00:02:44.778
118
  当另外两组特征已���存在时,加入 R 和 S,
119
 
120
- 00:02:44.778 --> 00:02:48.307
121
  分别平均增加四点八零和三点零一个百分点。
122
 
123
- 00:02:48.547 --> 00:02:52.390
124
  加入 I 的平均变化为负零点二七,区间跨过零。
125
 
126
- 00:02:52.630 --> 00:02:56.893
127
  结果支持补充缺失线索,而不是不断增加重复表征。
128
 
129
- 00:02:58.217 --> 00:03:01.034
130
  WristHARBench 提供基准数据;
131
 
132
- 00:03:01.034 --> 00:03:04.656
133
  WISP 提供方法代码、保存结果和实际保留的搜索模型。
134
 
135
- 00:03:04.896 --> 00:03:09.297
136
  三百六十个检查点不等于全部模型,也不代表完整复测。
137
 
138
- 00:03:09.537 --> 00:03:11.598
139
  从公开入口,继续探索。
 
1
  WEBVTT
2
 
3
+ 00:00:00.650 --> 00:00:07.868
4
  同样是手腕运动,走路、做饭和整理物品,留下的线索并不相同。
5
 
6
+ 00:00:07.868 --> 00:00:12.396
7
  局部形状、节律、先后次序,还有不同轴之间的关系,
8
 
9
+ 00:00:12.396 --> 00:00:13.796
10
  都可能帮助识别活动。
11
 
12
+ 00:00:15.015 --> 00:00:18.725
13
+ 如果把身体不同部位的传感器混在一起,
14
 
15
+ 00:00:18.725 --> 00:00:23.541
16
+ 结果就难以回答:只看手腕,究竟能识别什么?
17
 
18
+ 00:00:23.541 --> 00:00:28.543
19
  WristHARBench 整理了九个腕部三轴加速度数据集。
20
 
21
+ 00:00:28.543 --> 00:00:31.560
22
  另有两个多通道惯性数据集,单独分析。
23
 
24
+ 00:00:32.721 --> 00:00:38.231
25
  每次划分中,不同参与者进入训练、验证和测试组,
26
 
27
+ 00:00:38.231 --> 00:00:40.885
28
  同一个人不会跨组出现。
29
 
30
+ 00:00:40.885 --> 00:00:45.776
31
  模型设置只用验证组选择,测试组留到最后。
32
 
33
+ 00:00:45.776 --> 00:00:49.406
34
  所有方法使用相同的窗口和划分,重复五次。
35
 
36
+ 00:00:50.709 --> 00:00:56.600
37
  WISP 把同一段信号转换为多种表征。
38
 
39
+ 00:00:56.600 --> 00:01:00.837
40
  统计和频谱描述整体变化与节律;
41
 
42
+ 00:01:00.837 --> 00:01:06.764
43
  其他分支保留局部形状、相对次序和轴间关系。
44
 
45
+ 00:01:06.764 --> 00:01:10.657
46
  这里的曲面来自合成信号,是机制示意。
47
 
48
+ 00:01:11.865 --> 00:01:18.138
49
  随机卷积核沿时间移动,对不同的局部波形产生响应。
50
 
51
+ 00:01:18.138 --> 00:01:23.120
52
  选定一条截线,就能看到同一窗口内的响应变化。
53
 
54
+ 00:01:23.120 --> 00:01:29.938
55
  正值比例、最大值、均值和标准差,把整条响应汇总为四个特征。
56
 
57
+ 00:01:31.146 --> 00:01:38.964
58
  状态与次序表征,把窗口内的数值按分位数转成相对状态。
59
 
60
+ 00:01:38.964 --> 00:01:44.046
61
  再统计各状态出现多少,以及它们如何相互转移。
62
 
63
+ 00:01:44.046 --> 00:01:48.917
64
  矩阵记录的是转移概率,不是注意力权重。
65
 
66
+ 00:01:50.104 --> 00:01:53.231
67
+ 几何、区间和形状表征,
68
 
69
+ 00:01:53.231 --> 00:01:58.849
70
+ 同时读取轴间结构、局部统计和子序列匹配。
71
+
72
+ 00:01:58.849 --> 00:02:03.940
73
  模板沿窗口移动,距离最低的位置给出最佳匹配。
74
 
75
+ 00:02:03.940 --> 00:02:07.558
76
  保留下来的线索,包括匹配质量和出现位置。
77
 
78
+ 00:02:08.735 --> 00:02:12.590
79
  互补线索可以在特征层组合。
80
 
81
+ 00:02:12.590 --> 00:02:16.844
82
  CIS 把卷积特征 R、区间特征 I,
83
 
84
+ 00:02:16.844 --> 00:02:21.826
85
  以及频谱、时间和轴间特征 S,拼接成一个向量。
86
 
87
+ 00:02:21.826 --> 00:02:25.572
88
  随后由同一个分类器读取,输出各活动的分数。
89
 
90
+ 00:02:26.918 --> 00:02:31.791
91
  另一种组合发生在分数层:各成员先独立预测,
92
 
93
+ 00:02:31.791 --> 00:02:34.882
94
  再按相同的活动顺序对齐。
95
 
96
+ 00:02:34.882 --> 00:02:40.046
97
  CSE 等权平均三个成员,ESE 使用加权平均。
98
 
99
+ 00:02:40.046 --> 00:02:42.936
100
  这与特征拼接,是两种不同的组合方式。
101
 
102
+ 00:02:44.409 --> 00:02:46.864
103
  在冻结的论文评估范围内,
104
 
105
+ 00:02:46.864 --> 00:02:53.421
106
  随机搜索版 WISP 的平均 Macro F1 为百分之七十点二零。
107
 
108
+ 00:02:53.421 --> 00:02:58.573
109
  十八个外部基线中,最高的 RDST 为百分之五十九点六二,
110
 
111
+ 00:02:58.573 --> 00:03:01.955
112
  相差十点五八个百分点。
113
 
114
+ 00:03:01.955 --> 00:03:06.155
115
  这是九个数据集的平均结果,不代表每个数据集都获胜。
116
 
117
+ 00:03:07.419 --> 00:03:11.110
118
  受控消融进一步检验互补性。
119
 
120
+ 00:03:11.110 --> 00:03:15.256
121
  当另外两组特征已���存在时,加入 R 和 S,
122
 
123
+ 00:03:15.256 --> 00:03:20.419
124
  分别平均增加四点八零和三点零一个百分点。
125
 
126
+ 00:03:20.419 --> 00:03:25.801
127
  加入 I 的平均变化为负零点二七,区间跨过零。
128
 
129
+ 00:03:25.801 --> 00:03:30.365
130
  结果支持补充缺失线索,而不是不断增加重复表征。
131
 
132
+ 00:03:31.505 --> 00:03:34.396
133
  WristHARBench 提供基准数据;
134
 
135
+ 00:03:34.396 --> 00:03:39.719
136
  WISP 提供方法代码、保存结果和实际保留的搜索模型。
137
 
138
+ 00:03:39.719 --> 00:03:44.778
139
  三百六十个检查点不等于全部模型,也不代表完整复测。
140
 
141
+ 00:03:44.778 --> 00:03:46.542
142
  从公开入口,继续探索。
videos/WISP_zh_mimo_voice.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a06a03c57cfc7e2cd6936fd232a7e4ee931255250eb08a4a16eff32635a1f8c1
3
+ size 105702723
videos/WISP_zh_mimo_voice.wav ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:3b9ef2e4d01bc0527ecc2ac2edfc5377c0a261b05e8d39e08ed9fd6d8e4f5726
3
+ size 21804844
videos/WISP_zh_script.txt CHANGED
@@ -1,69 +1,59 @@
1
  00:00:00,000 从腕部信号,读懂活动
2
  同样是手腕运动,走路、做饭和整理物品,留下的线索并不相同。
3
- 局部形状、节律、先后次序,还有不同轴之间的关系,
4
- 都可能帮助识别活动。
5
 
6
- 00:00:13,100 WristHARBench:统一比较
7
- 如果把身体不同部位的传感器混在一起,结果就难以回答:只看手腕,
8
- 究竟能识别什么?
9
  WristHARBench 整理了九个腕部三轴加速度数据集。
10
  另有两个多通道惯性数据集,单独分析。
11
 
12
- 00:00:30,100 先按人划分,再比较模型
13
- 每次划分中,不同参与者进入训练、验证和测试组,
14
- 同一个人不会跨组出现。
15
  模型设置只用验证组选择,测试组留到最后。
16
  所有方法使用相同的窗口和划分,重复五次。
17
 
18
- 00:00:46,100 同一窗口,多种读法
19
  WISP 把同一段信号转换为多种表征。
20
- 统计和频谱描述整体变化与节律;
21
- 其他分支保留局部形状、相对次序和轴间关系。
22
  这里的曲面来自合成信号,是机制示意。
23
 
24
- 00:01:00,600 RC:响应如何变成特征
25
  随机卷积核沿时间移动,对不同的局部波形产生响应。
26
  选定一条截线,就能看到同一窗口内的响应变化。
27
  正值比例、最大值、均值和标准差,把整条响应汇总为四个特征。
28
 
29
- 00:01:18,600 SO:保留相对状态与次序
30
  状态与次序表征,把窗口内的数值按分位数转成相对状态。
31
  再统计各状态出现多少,以及它们如何相互转移。
32
  矩阵记录的是转移概率,不是注意力权重。
33
 
34
- 00:01:32,600 GIS:在窗口里寻找形状
35
  几何、区间和形状表征,同时读取轴间结构、局部统计和子序列匹配。
36
  模板沿窗口移动,距离最低的位置给出最佳匹配。
37
  保留下来的线索,包括匹配质量和出现位置。
38
 
39
- 00:01:47,433 CIS:先拼接特征
40
  互补线索可以在特征层组合。
41
- CIS 把卷积特征 R、区间特征 I,
42
- 以及频谱、时间和轴间特征 S,拼接成一个向量。
43
  随后由同一个分类器读取,输出各活动的分数。
44
 
45
- 00:02:03,433 CSE / ESE:再融合分数
46
- 另一种组合发生在分数层:各成员先独立预测,
47
- 再按相同的活动顺序对齐。
48
  CSE 等权平均三个成员,ESE 使用加权平均。
49
  这与特征拼接,是两种不同的组合方式。
50
 
51
- 00:02:18,767 九个数据集,等权平均
52
- 在冻结的论文评估范围内,
53
- 随机搜索版 WISP 的平均 Macro F1 为百分之七十点二零。
54
- 十八个外部基线中,最高的 RDST 为百分之五十九点六二,
55
- 相差十点五八个百分点。
56
  这是九个数据集的平均结果,不代表每个数据集都获胜。
57
 
58
- 00:02:37,933 增加什么信息,比增加多少更关键
59
  受控消融进一步检验互补性。
60
- 当另外两组特征已经存在时,加入 R 和 S,
61
- 分别平均增加四点八零和三点零一个百分点。
62
  加入 I 的平均变化为负零点二七,区间跨过零。
63
  结果支持补充缺失线索,而不是不断增加重复表征。
64
 
65
- 00:02:57,767 从统一评估,到可组合表征
66
- WristHARBench 提供基准数据;
67
- WISP 提供方法代码、保存结果和实际保留的搜索模型。
68
  三百六十个检查点不等于全部模型,也不代表完整复测。
69
- 从公开入口,继续探索。
 
1
  00:00:00,000 从腕部信号,读懂活动
2
  同样是手腕运动,走路、做饭和整理物品,留下的线索并不相同。
3
+ 局部形状、节律、先后次序,还有不同轴之间的关系,都可能帮助识别活动。
 
4
 
5
+ 00:00:14,400 WristHARBench:统一比较
6
+ 如果把身体不同部位的传感器混在一起,结果就难以回答:只看手腕,究竟能识别什么?
 
7
  WristHARBench 整理了九个腕部三轴加速度数据集。
8
  另有两个多通道惯性数据集,单独分析。
9
 
10
+ 00:00:32,133 先按人划分,再比较模型
11
+ 每次划分中,不同参与者进入训练、验证和测试组,同一个人不会跨组出现。
 
12
  模型设置只用验证组选择,测试组留到最后。
13
  所有方法使用相同的窗口和划分,重复五次。
14
 
15
+ 00:00:49,967 同一窗口,多种读法
16
  WISP 把同一段信号转换为多种表征。
17
+ 统计和频谱描述整体变化与节律;其他分支保留局部形状、相对次序和轴间关系。
 
18
  这里的曲面来自合成信号,是机制示意。
19
 
20
+ 00:01:11,233 RC:响应如何变成特征
21
  随机卷积核沿时间移动,对不同的局部波形产生响应。
22
  选定一条截线,就能看到同一窗口内的响应变化。
23
  正值比例、最大值、均值和标准差,把整条响应汇总为四个特征。
24
 
25
+ 00:01:30,567 SO:保留相对状态与次序
26
  状态与次序表征,把窗口内的数值按分位数转成相对状态。
27
  再统计各状态出现多少,以及它们如何相互转移。
28
  矩阵记录的是转移概率,不是注意力权重。
29
 
30
+ 00:01:49,500 GIS:在窗口里寻找形状
31
  几何、区间和形状表征,同时读取轴间结构、局部统计和子序列匹配。
32
  模板沿窗口移动,距离最低的位置给出最佳匹配。
33
  保留下来的线索,包括匹配质量和出现位置。
34
 
35
+ 00:02:08,167 CIS:先拼接特征
36
  互补线索可以在特征层组合。
37
+ CIS 把卷积特征 R、区间特征 I,以及频谱、时间和轴间特征 S,拼接成一个向量。
 
38
  随后由同一个分类器读取,输出各活动的分数。
39
 
40
+ 00:02:26,267 CSE / ESE:再融合分数
41
+ 另一种组合发生在分数层:各成员先独立预测,再按相同的活动顺序对齐。
 
42
  CSE 等权平均三个成员,ESE 使用加权平均。
43
  这与特征拼接,是两种不同的组合方式。
44
 
45
+ 00:02:43,767 九个数据集,等权平均
46
+ 在冻结的论文评估范围内,随机搜索版 WISP 的平均 Macro F1 为百分之七十点二零。
47
+ 十八个外部基线中,最高的 RDST 为百分之五十九点六二,相差十点五八个百分点。
 
 
48
  这是九个数据集的平均结果,不代表每个数据集都获胜。
49
 
50
+ 00:03:06,733 增加什么信息,比增加多少更关键
51
  受控消融进一步检验互补性。
52
+ 当另外两组特征已经存在时,加入 R 和 S,分别平均增加四点八零和三点零一个百分点。
 
53
  加入 I 的平均变化为负零点二七,区间跨过零。
54
  结果支持补充缺失线索,而不是不断增加重复表征。
55
 
56
+ 00:03:30,967 从统一评估,到可组合表征
57
+ WristHARBench 提供基准数据;WISP 提供方法代码、保存结果和实际保留的搜索模型。
 
58
  三百六十个检查点不等于全部模型,也不代表完整复测。
59
+ 从公开入口,继续探索。
videos/WISP_zh_system_voice.mp4 CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:badb93c00ee41539777037c4131a5e6a2ab82822e8130795987182e26e560d95
3
- size 86822326
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a06a03c57cfc7e2cd6936fd232a7e4ee931255250eb08a4a16eff32635a1f8c1
3
+ size 105702723
videos/archive/20261023/WISP_en.srt ADDED
@@ -0,0 +1,235 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 1
2
+ 00:00:00,450 --> 00:00:02,424
3
+ Walking, cooking, and handling objects
4
+
5
+ 2
6
+ 00:00:02,424 --> 00:00:04,398
7
+ leave different clues in wrist motion.
8
+
9
+ 3
10
+ 00:00:04,638 --> 00:00:07,042
11
+ Local shape, rhythm, temporal order, and
12
+
13
+ 4
14
+ 00:00:07,042 --> 00:00:10,709
15
+ relationships between axes can all help identify an activity.
16
+
17
+ 5
18
+ 00:00:13,450 --> 00:00:15,595
19
+ When benchmarks mix sensor locations, they
20
+
21
+ 6
22
+ 00:00:15,595 --> 00:00:17,843
23
+ cannot isolate what the wrist alone reveals.
24
+
25
+ 7
26
+ 00:00:18,083 --> 00:00:20,178
27
+ WristHARBench brings together nine wrist
28
+
29
+ 8
30
+ 00:00:20,178 --> 00:00:22,902
31
+ acceleration datasets under one evaluation contract.
32
+
33
+ 9
34
+ 00:00:23,142 --> 00:00:25,031
35
+ Two additional multichannel inertial
36
+
37
+ 10
38
+ 00:00:25,031 --> 00:00:27,602
39
+ datasets support a separate sensitivity analysis.
40
+
41
+ 11
42
+ 00:00:30,450 --> 00:00:32,538
43
+ Within each repeat, different people
44
+
45
+ 12
46
+ 00:00:32,538 --> 00:00:35,091
47
+ enter training, validation, and test groups.
48
+
49
+ 13
50
+ 00:00:35,331 --> 00:00:37,231
51
+ No person appears in two groups.
52
+
53
+ 14
54
+ 00:00:37,231 --> 00:00:40,259
55
+ Validation guides model choice; testing comes last.
56
+
57
+ 15
58
+ 00:00:40,499 --> 00:00:44,406
59
+ Methods share the same windows and splits, across five repeated holdouts.
60
+
61
+ 16
62
+ 00:00:46,450 --> 00:00:50,120
63
+ WISP reads the same signal window through complementary representations.
64
+
65
+ 17
66
+ 00:00:50,360 --> 00:00:53,740
67
+ Statistics and spectra describe overall variation and rhythm;
68
+
69
+ 18
70
+ 00:00:53,740 --> 00:00:57,121
71
+ other branches retain shape, order, and cross-axis structure.
72
+
73
+ 19
74
+ 00:00:57,361 --> 00:01:00,377
75
+ These landscapes illustrate operators on synthetic signals.
76
+
77
+ 20
78
+ 00:01:01,683 --> 00:01:03,816
79
+ Random convolution kernels move through
80
+
81
+ 21
82
+ 00:01:03,816 --> 00:01:06,223
83
+ time, responding to local waveform patterns.
84
+
85
+ 22
86
+ 00:01:06,463 --> 00:01:09,757
87
+ A selected slice reveals the responses within one signal window.
88
+
89
+ 23
90
+ 00:01:09,997 --> 00:01:12,797
91
+ Positive proportion, maximum, mean, and standard
92
+
93
+ 24
94
+ 00:01:12,797 --> 00:01:15,597
95
+ deviation pool that response into four features.
96
+
97
+ 25
98
+ 00:01:19,683 --> 00:01:21,788
99
+ The state-and-order representation maps
100
+
101
+ 26
102
+ 00:01:21,788 --> 00:01:24,487
103
+ values to relative, within-window quantile states.
104
+
105
+ 27
106
+ 00:01:24,727 --> 00:01:27,870
107
+ It counts how often states occur and how they transition to each other.
108
+
109
+ 28
110
+ 00:01:28,110 --> 00:01:31,806
111
+ The matrix contains transition probabilities, not attention weights.
112
+
113
+ 29
114
+ 00:01:33,683 --> 00:01:36,487
115
+ Geometry, interval, and shape features combine
116
+
117
+ 30
118
+ 00:01:36,487 --> 00:01:40,389
119
+ cross-axis structure, local statistics, and subsequence matches.
120
+
121
+ 31
122
+ 00:01:40,629 --> 00:01:42,633
123
+ A template moves along the window. The
124
+
125
+ 32
126
+ 00:01:42,633 --> 00:01:44,847
127
+ lowest distance identifies its best match.
128
+
129
+ 33
130
+ 00:01:45,087 --> 00:01:47,587
131
+ Both match quality and location become features.
132
+
133
+ 34
134
+ 00:01:48,883 --> 00:01:51,624
135
+ Complementary clues can be combined at the feature level.
136
+
137
+ 35
138
+ 00:01:51,864 --> 00:01:55,855
139
+ features I, and spectral, temporal, and cross-axis features S.
140
+
141
+ 36
142
+ 00:01:55,855 --> 00:01:59,073
143
+ features I, and spectral, temporal, and cross-axis
144
+
145
+ 37
146
+ 00:01:59,313 --> 00:02:03,210
147
+ A single classifier reads the combined vector and produces activity scores.
148
+
149
+ 38
150
+ 00:02:04,883 --> 00:02:07,104
151
+ Score ensembles combine independently fitted
152
+
153
+ 39
154
+ 00:02:07,104 --> 00:02:09,425
155
+ members after aligning their activity classes.
156
+
157
+ 40
158
+ 00:02:09,665 --> 00:02:13,544
159
+ CSE averages three members equally; ESE uses weighted averaging.
160
+
161
+ 41
162
+ 00:02:13,784 --> 00:02:17,308
163
+ This is different from concatenating features before one classifier.
164
+
165
+ 42
166
+ 00:02:19,883 --> 00:02:22,641
167
+ On the frozen paper evaluation, WISP with
168
+
169
+ 43
170
+ 00:02:22,641 --> 00:02:26,004
171
+ random search reaches 70.20 percent mean macro F1.
172
+
173
+ 44
174
+ 00:02:26,244 --> 00:02:29,088
175
+ RDST, the strongest of eighteen external
176
+
177
+ 45
178
+ 00:02:29,088 --> 00:02:32,571
179
+ baselines on this average, reaches 59.62 percent.
180
+
181
+ 46
182
+ 00:02:32,811 --> 00:02:35,657
183
+ The difference is 10.58 percentage points. This
184
+
185
+ 47
186
+ 00:02:35,657 --> 00:02:38,623
187
+ is a dataset average, not a win on every dataset.
188
+
189
+ 48
190
+ 00:02:39,917 --> 00:02:43,940
191
+ Controlled ablations test complementary information with a fixed classifier.
192
+
193
+ 49
194
+ 00:02:44,180 --> 00:02:47,181
195
+ With the other two groups already present,
196
+
197
+ 50
198
+ 00:02:47,181 --> 00:02:50,610
199
+ adding R gains 4.80 points; adding S gains 3.01.
200
+
201
+ 51
202
+ 00:02:50,850 --> 00:02:53,367
203
+ Adding I changes the mean by minus 0.27
204
+
205
+ 52
206
+ 00:02:53,367 --> 00:02:55,883
207
+ points, with an interval crossing zero.
208
+
209
+ 53
210
+ 00:02:56,123 --> 00:02:58,322
211
+ The evidence supports adding missing clues,
212
+
213
+ 54
214
+ 00:02:58,322 --> 00:03:00,623
215
+ rather than more overlapping representations.
216
+
217
+ 55
218
+ 00:03:01,917 --> 00:03:04,910
219
+ WristHARBench provides benchmark data. WISP provides
220
+
221
+ 56
222
+ 00:03:04,910 --> 00:03:08,076
223
+ method code, saved results, and retained search models.
224
+
225
+ 57
226
+ 00:03:08,316 --> 00:03:10,712
227
+ The 360 retained checkpoints do not cover
228
+
229
+ 58
230
+ 00:03:10,712 --> 00:03:13,516
231
+ every method or imply complete inference reruns.
232
+
233
+ 59
234
+ 00:03:13,756 --> 00:03:15,325
235
+ Explore the public releases.
videos/archive/20261023/WISP_en.vtt ADDED
@@ -0,0 +1,178 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ WEBVTT
2
+
3
+ 00:00:00.450 --> 00:00:02.424
4
+ Walking, cooking, and handling objects
5
+
6
+ 00:00:02.424 --> 00:00:04.398
7
+ leave different clues in wrist motion.
8
+
9
+ 00:00:04.638 --> 00:00:07.042
10
+ Local shape, rhythm, temporal order, and
11
+
12
+ 00:00:07.042 --> 00:00:10.709
13
+ relationships between axes can all help identify an activity.
14
+
15
+ 00:00:13.450 --> 00:00:15.595
16
+ When benchmarks mix sensor locations, they
17
+
18
+ 00:00:15.595 --> 00:00:17.843
19
+ cannot isolate what the wrist alone reveals.
20
+
21
+ 00:00:18.083 --> 00:00:20.178
22
+ WristHARBench brings together nine wrist
23
+
24
+ 00:00:20.178 --> 00:00:22.902
25
+ acceleration datasets under one evaluation contract.
26
+
27
+ 00:00:23.142 --> 00:00:25.031
28
+ Two additional multichannel inertial
29
+
30
+ 00:00:25.031 --> 00:00:27.602
31
+ datasets support a separate sensitivity analysis.
32
+
33
+ 00:00:30.450 --> 00:00:32.538
34
+ Within each repeat, different people
35
+
36
+ 00:00:32.538 --> 00:00:35.091
37
+ enter training, validation, and test groups.
38
+
39
+ 00:00:35.331 --> 00:00:37.231
40
+ No person appears in two groups.
41
+
42
+ 00:00:37.231 --> 00:00:40.259
43
+ Validation guides model choice; testing comes last.
44
+
45
+ 00:00:40.499 --> 00:00:44.406
46
+ Methods share the same windows and splits, across five repeated holdouts.
47
+
48
+ 00:00:46.450 --> 00:00:50.120
49
+ WISP reads the same signal window through complementary representations.
50
+
51
+ 00:00:50.360 --> 00:00:53.740
52
+ Statistics and spectra describe overall variation and rhythm;
53
+
54
+ 00:00:53.740 --> 00:00:57.121
55
+ other branches retain shape, order, and cross-axis structure.
56
+
57
+ 00:00:57.361 --> 00:01:00.377
58
+ These landscapes illustrate operators on synthetic signals.
59
+
60
+ 00:01:01.683 --> 00:01:03.816
61
+ Random convolution kernels move through
62
+
63
+ 00:01:03.816 --> 00:01:06.223
64
+ time, responding to local waveform patterns.
65
+
66
+ 00:01:06.463 --> 00:01:09.757
67
+ A selected slice reveals the responses within one signal window.
68
+
69
+ 00:01:09.997 --> 00:01:12.797
70
+ Positive proportion, maximum, mean, and standard
71
+
72
+ 00:01:12.797 --> 00:01:15.597
73
+ deviation pool that response into four features.
74
+
75
+ 00:01:19.683 --> 00:01:21.788
76
+ The state-and-order representation maps
77
+
78
+ 00:01:21.788 --> 00:01:24.487
79
+ values to relative, within-window quantile states.
80
+
81
+ 00:01:24.727 --> 00:01:27.870
82
+ It counts how often states occur and how they transition to each other.
83
+
84
+ 00:01:28.110 --> 00:01:31.806
85
+ The matrix contains transition probabilities, not attention weights.
86
+
87
+ 00:01:33.683 --> 00:01:36.487
88
+ Geometry, interval, and shape features combine
89
+
90
+ 00:01:36.487 --> 00:01:40.389
91
+ cross-axis structure, local statistics, and subsequence matches.
92
+
93
+ 00:01:40.629 --> 00:01:42.633
94
+ A template moves along the window. The
95
+
96
+ 00:01:42.633 --> 00:01:44.847
97
+ lowest distance identifies its best match.
98
+
99
+ 00:01:45.087 --> 00:01:47.587
100
+ Both match quality and location become features.
101
+
102
+ 00:01:48.883 --> 00:01:51.624
103
+ Complementary clues can be combined at the feature level.
104
+
105
+ 00:01:51.864 --> 00:01:55.855
106
+ features I, and spectral, temporal, and cross-axis features S.
107
+
108
+ 00:01:55.855 --> 00:01:59.073
109
+ features I, and spectral, temporal, and cross-axis
110
+
111
+ 00:01:59.313 --> 00:02:03.210
112
+ A single classifier reads the combined vector and produces activity scores.
113
+
114
+ 00:02:04.883 --> 00:02:07.104
115
+ Score ensembles combine independently fitted
116
+
117
+ 00:02:07.104 --> 00:02:09.425
118
+ members after aligning their activity classes.
119
+
120
+ 00:02:09.665 --> 00:02:13.544
121
+ CSE averages three members equally; ESE uses weighted averaging.
122
+
123
+ 00:02:13.784 --> 00:02:17.308
124
+ This is different from concatenating features before one classifier.
125
+
126
+ 00:02:19.883 --> 00:02:22.641
127
+ On the frozen paper evaluation, WISP with
128
+
129
+ 00:02:22.641 --> 00:02:26.004
130
+ random search reaches 70.20 percent mean macro F1.
131
+
132
+ 00:02:26.244 --> 00:02:29.088
133
+ RDST, the strongest of eighteen external
134
+
135
+ 00:02:29.088 --> 00:02:32.571
136
+ baselines on this average, reaches 59.62 percent.
137
+
138
+ 00:02:32.811 --> 00:02:35.657
139
+ The difference is 10.58 percentage points. This
140
+
141
+ 00:02:35.657 --> 00:02:38.623
142
+ is a dataset average, not a win on every dataset.
143
+
144
+ 00:02:39.917 --> 00:02:43.940
145
+ Controlled ablations test complementary information with a fixed classifier.
146
+
147
+ 00:02:44.180 --> 00:02:47.181
148
+ With the other two groups already present,
149
+
150
+ 00:02:47.181 --> 00:02:50.610
151
+ adding R gains 4.80 points; adding S gains 3.01.
152
+
153
+ 00:02:50.850 --> 00:02:53.367
154
+ Adding I changes the mean by minus 0.27
155
+
156
+ 00:02:53.367 --> 00:02:55.883
157
+ points, with an interval crossing zero.
158
+
159
+ 00:02:56.123 --> 00:02:58.322
160
+ The evidence supports adding missing clues,
161
+
162
+ 00:02:58.322 --> 00:03:00.623
163
+ rather than more overlapping representations.
164
+
165
+ 00:03:01.917 --> 00:03:04.910
166
+ WristHARBench provides benchmark data. WISP provides
167
+
168
+ 00:03:04.910 --> 00:03:08.076
169
+ method code, saved results, and retained search models.
170
+
171
+ 00:03:08.316 --> 00:03:10.712
172
+ The 360 retained checkpoints do not cover
173
+
174
+ 00:03:10.712 --> 00:03:13.516
175
+ every method or imply complete inference reruns.
176
+
177
+ 00:03:13.756 --> 00:03:15.325
178
+ Explore the public releases.
videos/archive/20261023/WISP_en_cover.png ADDED

Git LFS Details

  • SHA256: 15d7c187ce4de39131b93cdf7b3bbea7b91cf350b2be7007d5a23d5d02ea2cfe
  • Pointer size: 131 Bytes
  • Size of remote file: 138 kB
videos/archive/20261023/WISP_en_script.txt ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 00:00:00,000 Reading activity at the wrist
2
+ Walking, cooking, and handling objects
3
+ leave different clues in wrist motion.
4
+ Local shape, rhythm, temporal order, and
5
+ relationships between axes can all help identify an activity.
6
+
7
+ 00:00:13,000 WristHARBench: a common comparison
8
+ When benchmarks mix sensor locations, they
9
+ cannot isolate what the wrist alone reveals.
10
+ WristHARBench brings together nine wrist
11
+ acceleration datasets under one evaluation contract.
12
+ Two additional multichannel inertial
13
+ datasets support a separate sensitivity analysis.
14
+
15
+ 00:00:30,000 Separate people before comparing models
16
+ Within each repeat, different people
17
+ enter training, validation, and test groups.
18
+ No person appears in two groups.
19
+ Validation guides model choice; testing comes last.
20
+ Methods share the same windows and splits, across five repeated holdouts.
21
+
22
+ 00:00:46,000 One window, several readings
23
+ WISP reads the same signal window through complementary representations.
24
+ Statistics and spectra describe overall variation and rhythm;
25
+ other branches retain shape, order, and cross-axis structure.
26
+ These landscapes illustrate operators on synthetic signals.
27
+
28
+ 00:01:01,233 RC: from responses to features
29
+ Random convolution kernels move through
30
+ time, responding to local waveform patterns.
31
+ A selected slice reveals the responses within one signal window.
32
+ Positive proportion, maximum, mean, and standard
33
+ deviation pool that response into four features.
34
+
35
+ 00:01:19,233 SO: relative states and their order
36
+ The state-and-order representation maps
37
+ values to relative, within-window quantile states.
38
+ It counts how often states occur and how they transition to each other.
39
+ The matrix contains transition probabilities, not attention weights.
40
+
41
+ 00:01:33,233 GIS: finding shapes within a window
42
+ Geometry, interval, and shape features combine
43
+ cross-axis structure, local statistics, and subsequence matches.
44
+ A template moves along the window. The
45
+ lowest distance identifies its best match.
46
+ Both match quality and location become features.
47
+
48
+ 00:01:48,433 CIS: concatenate features first
49
+ Complementary clues can be combined at the feature level.
50
+ features I, and spectral, temporal, and cross-axis features S.
51
+ features I, and spectral, temporal, and cross-axis
52
+ A single classifier reads the combined vector and produces activity scores.
53
+
54
+ 00:02:04,433 CSE / ESE: combine class scores
55
+ Score ensembles combine independently fitted
56
+ members after aligning their activity classes.
57
+ CSE averages three members equally; ESE uses weighted averaging.
58
+ This is different from concatenating features before one classifier.
59
+
60
+ 00:02:19,433 Nine datasets, equal weight
61
+ On the frozen paper evaluation, WISP with
62
+ random search reaches 70.20 percent mean macro F1.
63
+ RDST, the strongest of eighteen external
64
+ baselines on this average, reaches 59.62 percent.
65
+ The difference is 10.58 percentage points. This
66
+ is a dataset average, not a win on every dataset.
67
+
68
+ 00:02:39,467 What information does a branch add?
69
+ Controlled ablations test complementary information with a fixed classifier.
70
+ With the other two groups already present,
71
+ adding R gains 4.80 points; adding S gains 3.01.
72
+ Adding I changes the mean by minus 0.27
73
+ points, with an interval crossing zero.
74
+ The evidence supports adding missing clues,
75
+ rather than more overlapping representations.
76
+
77
+ 00:03:01,467 From shared evaluation to composable representations
78
+ WristHARBench provides benchmark data. WISP provides
79
+ method code, saved results, and retained search models.
80
+ The 360 retained checkpoints do not cover
81
+ every method or imply complete inference reruns.
82
+ Explore the public releases.
videos/archive/20261023/WISP_en_system_voice.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b1d36676464b1eecd714802f6be25c770074b682f510a2142db2b55b7e4822c8
3
+ size 90249749
videos/archive/20261023/WISP_zh.srt ADDED
@@ -0,0 +1,183 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 1
2
+ 00:00:00,450 --> 00:00:05,905
3
+ 同样是手腕运动,走路、做饭和整理物品,留下的线索并不相同。
4
+
5
+ 2
6
+ 00:00:06,145 --> 00:00:10,439
7
+ 局部形状、节律、先后次序,还有不同轴之间的关系,
8
+
9
+ 3
10
+ 00:00:10,439 --> 00:00:12,229
11
+ 都可能帮助识别活动。
12
+
13
+ 4
14
+ 00:00:13,550 --> 00:00:19,077
15
+ 如果把身体不同部位的传感器混在一起,结果就难以回答:只看手腕,
16
+
17
+ 5
18
+ 00:00:19,077 --> 00:00:20,503
19
+ 究竟能识别什么?
20
+
21
+ 6
22
+ 00:00:20,743 --> 00:00:24,161
23
+ WristHARBench 整理了九个腕部三轴加速度数据集。
24
+
25
+ 7
26
+ 00:00:24,401 --> 00:00:27,664
27
+ 另有两个多通道惯性数据集,单独分析。
28
+
29
+ 8
30
+ 00:00:30,550 --> 00:00:34,744
31
+ 每次划分中,不同参与者进入训练、验证和测试组,
32
+
33
+ 9
34
+ 00:00:34,744 --> 00:00:36,750
35
+ 同一个人不会跨组出现。
36
+
37
+ 10
38
+ 00:00:36,990 --> 00:00:40,532
39
+ 模型设置只用验证组选择,测试组留到最后。
40
+
41
+ 11
42
+ 00:00:40,772 --> 00:00:44,503
43
+ 所有方法使用相同的窗口和划分,重复五次。
44
+
45
+ 12
46
+ 00:00:46,550 --> 00:00:49,420
47
+ WISP 把同一段信号转换为多种表征。
48
+
49
+ 13
50
+ 00:00:49,660 --> 00:00:52,435
51
+ 统计和频谱描述整体变化与节律;
52
+
53
+ 14
54
+ 00:00:52,435 --> 00:00:56,319
55
+ 其他分支保留局部形状、相对次序和轴间关系。
56
+
57
+ 15
58
+ 00:00:56,559 --> 00:00:59,751
59
+ 这里的曲面来自合成信号,是机制示意。
60
+
61
+ 16
62
+ 00:01:01,050 --> 00:01:05,351
63
+ 随机卷积核沿时间移动,对不同的局部波形产生响应。
64
+
65
+ 17
66
+ 00:01:05,591 --> 00:01:09,659
67
+ 选定一条截线,就能看到同一窗口内的响应变化。
68
+
69
+ 18
70
+ 00:01:09,899 --> 00:01:15,345
71
+ 正值比例、最大值、均值和标准差,把整条响应汇总为四个特征。
72
+
73
+ 19
74
+ 00:01:19,050 --> 00:01:23,868
75
+ 状态与次序表征,把窗口内的数值按分位数转成相对状态。
76
+
77
+ 20
78
+ 00:01:24,108 --> 00:01:28,133
79
+ 再统计各状态出现多少,以及它们如何相互转移。
80
+
81
+ 21
82
+ 00:01:28,373 --> 00:01:31,379
83
+ 矩阵记录的是转移概率,不是注意力权重。
84
+
85
+ 22
86
+ 00:01:33,050 --> 00:01:38,900
87
+ 几何、区间和形状表征,同时读取轴间结构、局部统计和子序列匹配。
88
+
89
+ 23
90
+ 00:01:39,140 --> 00:01:42,769
91
+ 模板沿窗口移动,距离最低的位置给出最佳匹配。
92
+
93
+ 24
94
+ 00:01:43,009 --> 00:01:46,571
95
+ 保留下来的线索,包括匹配质量和出现位置。
96
+
97
+ 25
98
+ 00:01:47,883 --> 00:01:50,102
99
+ 互补线索可以在特征层组合。
100
+
101
+ 26
102
+ 00:01:50,342 --> 00:01:53,625
103
+ CIS 把卷积特征 R、区间特征 I,
104
+
105
+ 27
106
+ 00:01:53,625 --> 00:01:57,600
107
+ 以及频谱、时间和轴间特征 S,拼接成一个向量。
108
+
109
+ 28
110
+ 00:01:57,840 --> 00:02:01,451
111
+ 随后由同一个分类器读取,输出各活动的分数。
112
+
113
+ 29
114
+ 00:02:03,883 --> 00:02:07,747
115
+ 另一种组合发生在分数层:各成员先独立预测,
116
+
117
+ 30
118
+ 00:02:07,747 --> 00:02:09,955
119
+ 再按相同的活动顺序对齐。
120
+
121
+ 31
122
+ 00:02:10,195 --> 00:02:14,525
123
+ CSE 等权平均三个成员,ESE 使用加权平均。
124
+
125
+ 32
126
+ 00:02:14,765 --> 00:02:17,912
127
+ 这与特征拼接,是两种不同的组合方式。
128
+
129
+ 33
130
+ 00:02:19,217 --> 00:02:20,903
131
+ 在冻结的论文评估范围内,
132
+
133
+ 34
134
+ 00:02:20,903 --> 00:02:25,682
135
+ 随机搜索版 WISP 的平均 Macro F1 为百分之七十点二零。
136
+
137
+ 35
138
+ 00:02:25,922 --> 00:02:30,753
139
+ 十八个外部基线中,最高的 RDST 为百分之五十九点六二,
140
+
141
+ 36
142
+ 00:02:30,753 --> 00:02:32,586
143
+ 相差十点五八个百分点。
144
+
145
+ 37
146
+ 00:02:32,826 --> 00:02:37,090
147
+ 这是九个数据集的平均结果,不代表每个数据集都获胜。
148
+
149
+ 38
150
+ 00:02:38,383 --> 00:02:40,657
151
+ 受控消融进一步检验互补性。
152
+
153
+ 39
154
+ 00:02:40,897 --> 00:02:44,778
155
+ 当另外两组特征已经存在时,加入 R 和 S,
156
+
157
+ 40
158
+ 00:02:44,778 --> 00:02:48,307
159
+ 分别平均增加四点八零和三点零一个百分点。
160
+
161
+ 41
162
+ 00:02:48,547 --> 00:02:52,390
163
+ 加入 I 的平均变化为负零点二七,区间跨过零。
164
+
165
+ 42
166
+ 00:02:52,630 --> 00:02:56,893
167
+ 结果支持补充缺失线索,而不是不断增加重复表征。
168
+
169
+ 43
170
+ 00:02:58,217 --> 00:03:01,034
171
+ WristHARBench 提供基准数据;
172
+
173
+ 44
174
+ 00:03:01,034 --> 00:03:04,656
175
+ WISP 提供方法代码、保存结果和实际保留的搜索模型。
176
+
177
+ 45
178
+ 00:03:04,896 --> 00:03:09,297
179
+ 三百六十个检查点不等于全部模型,也不代表完整复测。
180
+
181
+ 46
182
+ 00:03:09,537 --> 00:03:11,598
183
+ 从公开入口,继续探索。
videos/archive/20261023/WISP_zh.vtt ADDED
@@ -0,0 +1,139 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ WEBVTT
2
+
3
+ 00:00:00.450 --> 00:00:05.905
4
+ 同样是手腕运动,走路、做饭和整理物品,留下的线索并不相同。
5
+
6
+ 00:00:06.145 --> 00:00:10.439
7
+ 局部形状、节律、先后次序,还有不同轴之间的关系,
8
+
9
+ 00:00:10.439 --> 00:00:12.229
10
+ 都可能帮助识别活动。
11
+
12
+ 00:00:13.550 --> 00:00:19.077
13
+ 如果把身体不同部位的传感器混在一起,结果就难以回答:只看手腕,
14
+
15
+ 00:00:19.077 --> 00:00:20.503
16
+ 究竟能识别什么?
17
+
18
+ 00:00:20.743 --> 00:00:24.161
19
+ WristHARBench 整理了九个腕部三轴加速度数据集。
20
+
21
+ 00:00:24.401 --> 00:00:27.664
22
+ 另有两个多通道惯性数据集,单独分析。
23
+
24
+ 00:00:30.550 --> 00:00:34.744
25
+ 每次划分中,不同参与者进入训练、验证和测试组,
26
+
27
+ 00:00:34.744 --> 00:00:36.750
28
+ 同一个人不会跨组出现。
29
+
30
+ 00:00:36.990 --> 00:00:40.532
31
+ 模型设置只用验证组选择,测试组留到最后。
32
+
33
+ 00:00:40.772 --> 00:00:44.503
34
+ 所有方法使用相同的窗口和划分,重复五次。
35
+
36
+ 00:00:46.550 --> 00:00:49.420
37
+ WISP 把同一段信号转换为多种表征。
38
+
39
+ 00:00:49.660 --> 00:00:52.435
40
+ 统计和频谱描述整体变化与节律;
41
+
42
+ 00:00:52.435 --> 00:00:56.319
43
+ 其他分支保留局部形状、相对次序和轴间关系。
44
+
45
+ 00:00:56.559 --> 00:00:59.751
46
+ 这里的曲面来自合成信号,是机制示意。
47
+
48
+ 00:01:01.050 --> 00:01:05.351
49
+ 随机卷积核沿时间移动,对不同的局部波形产生响应。
50
+
51
+ 00:01:05.591 --> 00:01:09.659
52
+ 选定一条截线,就能看到同一窗口内的响应变化。
53
+
54
+ 00:01:09.899 --> 00:01:15.345
55
+ 正值比例、最大值、均值和标准差,把整条响应汇总为四个特征。
56
+
57
+ 00:01:19.050 --> 00:01:23.868
58
+ 状态与次序表征,把窗口内的数值按分位数转成相对状态。
59
+
60
+ 00:01:24.108 --> 00:01:28.133
61
+ 再统计各状态出现多少,以及它们如何相互转移。
62
+
63
+ 00:01:28.373 --> 00:01:31.379
64
+ 矩阵记录的是转移概率,不是注意力权重。
65
+
66
+ 00:01:33.050 --> 00:01:38.900
67
+ 几何、区间和形状表征,同时读取轴间结构、局部统计和子序列匹配。
68
+
69
+ 00:01:39.140 --> 00:01:42.769
70
+ 模板沿窗口移动,距离最低的位置给出最佳匹配。
71
+
72
+ 00:01:43.009 --> 00:01:46.571
73
+ 保留下来的线索,包括匹配质量和出现位置。
74
+
75
+ 00:01:47.883 --> 00:01:50.102
76
+ 互补线索可以在特征层组合。
77
+
78
+ 00:01:50.342 --> 00:01:53.625
79
+ CIS 把卷积特征 R、区间特征 I,
80
+
81
+ 00:01:53.625 --> 00:01:57.600
82
+ 以及频谱、时间和轴间特征 S,拼接成一个向量。
83
+
84
+ 00:01:57.840 --> 00:02:01.451
85
+ 随后由同一个分类器读取,输出各活动的分数。
86
+
87
+ 00:02:03.883 --> 00:02:07.747
88
+ 另一种组合发生在分数层:各成员先独立预测,
89
+
90
+ 00:02:07.747 --> 00:02:09.955
91
+ 再按相同的活动顺序对齐。
92
+
93
+ 00:02:10.195 --> 00:02:14.525
94
+ CSE 等权平均三个成员,ESE 使用加权平均。
95
+
96
+ 00:02:14.765 --> 00:02:17.912
97
+ 这与特征拼接,是两种不同的组合方式。
98
+
99
+ 00:02:19.217 --> 00:02:20.903
100
+ 在冻结的论文评估范围内,
101
+
102
+ 00:02:20.903 --> 00:02:25.682
103
+ 随机搜索版 WISP 的平均 Macro F1 为百分之七十点二零。
104
+
105
+ 00:02:25.922 --> 00:02:30.753
106
+ 十八个外部基线中,最高的 RDST 为百分之五十九点六二,
107
+
108
+ 00:02:30.753 --> 00:02:32.586
109
+ 相差十点五八个百分点。
110
+
111
+ 00:02:32.826 --> 00:02:37.090
112
+ 这是九个数据集的平均结果,不代表每个数据集都获胜。
113
+
114
+ 00:02:38.383 --> 00:02:40.657
115
+ 受控消融进一步检验互补性。
116
+
117
+ 00:02:40.897 --> 00:02:44.778
118
+ 当另外两组特征已经存在时,加入 R 和 S,
119
+
120
+ 00:02:44.778 --> 00:02:48.307
121
+ 分别平均增加四点八零和三点零一个百分点。
122
+
123
+ 00:02:48.547 --> 00:02:52.390
124
+ 加入 I 的平均变化为负零点二七,区间跨过零。
125
+
126
+ 00:02:52.630 --> 00:02:56.893
127
+ 结果支持补充缺失线索,而不是不断增加重复表征。
128
+
129
+ 00:02:58.217 --> 00:03:01.034
130
+ WristHARBench 提供基准数据;
131
+
132
+ 00:03:01.034 --> 00:03:04.656
133
+ WISP 提供方法代码、保存结果和实际保留的搜索模型。
134
+
135
+ 00:03:04.896 --> 00:03:09.297
136
+ 三百六十个检查点不等于全部模型,也不代表完整复测。
137
+
138
+ 00:03:09.537 --> 00:03:11.598
139
+ 从公开入口,继续探索。
videos/archive/20261023/WISP_zh_cover.png ADDED

Git LFS Details

  • SHA256: f7f33cdfcef368d2d9f484342dae5c7ab947dce60e1cf3f1890e63cab71c0775
  • Pointer size: 131 Bytes
  • Size of remote file: 131 kB
videos/archive/20261023/WISP_zh_script.txt ADDED
@@ -0,0 +1,69 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ 00:00:00,000 从腕部信号,读懂活动
2
+ 同样是手腕运动,走路、做饭和整理物品,留下的线索并不相同。
3
+ 局部形状、节律、先后次序,还有不同轴之间的关系,
4
+ 都可能帮助识别活动。
5
+
6
+ 00:00:13,100 WristHARBench:统一比较
7
+ 如果把身体不同部位的传感器混在一起,结果就难以回答:只看手腕,
8
+ 究竟能识别什么?
9
+ WristHARBench 整理了九个腕部三轴加速度数据集。
10
+ 另有两个多通道惯性数据集,单独分析。
11
+
12
+ 00:00:30,100 先按人划分,再比较模型
13
+ 每次划分中,不同参与者进入训练、验证和测试组,
14
+ 同一个人不会跨组出现。
15
+ 模型设置只用验证组选择,测试组留到最后。
16
+ 所有方法使用相同的窗口和划分,重复五次。
17
+
18
+ 00:00:46,100 同一窗口,多种读法
19
+ WISP 把同一段信号转换为多种表征。
20
+ 统计和频谱描述整体变化与节律;
21
+ 其他分支保留局部形状、相对次序和轴间关系。
22
+ 这里的曲面来自合成信号,是机制示意。
23
+
24
+ 00:01:00,600 RC:响应如何变成特征
25
+ 随机卷积核沿时间移动,对不同的局部波形产生响应。
26
+ 选定一条截线,就能看到同一窗口内的响应变化。
27
+ 正值比例、最大值、均值和标准差,把整条响应汇总为四个特征。
28
+
29
+ 00:01:18,600 SO:保留相对状态与次序
30
+ 状态与次序表征,把窗口内的数值按分位数转成相对状态。
31
+ 再统计各状态出现多少,以及它们如何相互转移。
32
+ 矩阵记录的是转移概率,不是注意力权重。
33
+
34
+ 00:01:32,600 GIS:在窗口里寻找形状
35
+ 几何、区间和形状表征,同时读取轴间结构、局部统计和子序列匹配。
36
+ 模板沿窗口移动,距离最低的位置给出最佳匹配。
37
+ 保留下来的线索,包括匹配质量和出现位置。
38
+
39
+ 00:01:47,433 CIS:先拼接特征
40
+ 互补线索可以在特征层组合。
41
+ CIS 把卷积特征 R、区间特征 I,
42
+ 以及频谱、时间和轴间特征 S,拼接成一个向量。
43
+ 随后由同一个分类器读取,输出各活动的分数。
44
+
45
+ 00:02:03,433 CSE / ESE:再融合分数
46
+ 另一种组合发生在分数层:各成员先独立预测,
47
+ 再按相同的活动顺序对齐。
48
+ CSE 等权平均三个成员,ESE 使用加权平均。
49
+ 这与特征拼接,是两种不同的组合方式。
50
+
51
+ 00:02:18,767 九个数据集,等权平均
52
+ 在冻结的论文评估范围内,
53
+ 随机搜索版 WISP 的平均 Macro F1 为百分之七十点二零。
54
+ 十八个外部基线中,最高的 RDST 为百分之五十九点六二,
55
+ 相差十点五八个百分点。
56
+ 这是九个数据集的平均结果,不代表每个数据集都获胜。
57
+
58
+ 00:02:37,933 增加什么信息,比增加多少更关键
59
+ 受控消融进一步检验互补性。
60
+ 当另外两组特征已经存在时,加入 R 和 S,
61
+ 分别平均增加四点八零和三点零一个百分点。
62
+ 加入 I 的平均变化为负零点二七,区间跨过零。
63
+ 结果支持补充缺失线索,而不是不断增加重复表征。
64
+
65
+ 00:02:57,767 从统一评估,到可组合表征
66
+ WristHARBench 提供基准数据;
67
+ WISP 提供方法代码、保存结果和实际保留的搜索模型。
68
+ 三百六十个检查点不等于全部模型,也不代表完整复测。
69
+ 从公开入口,继续探索。
videos/archive/20261023/WISP_zh_system_voice.mp4 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:badb93c00ee41539777037c4131a5e6a2ab82822e8130795987182e26e560d95
3
+ size 86822326
videos/release_manifest.json ADDED
@@ -0,0 +1,62 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "WISP_zh_mimo_voice.mp4": {
3
+ "bytes": 105702723,
4
+ "sha256": "a06a03c57cfc7e2cd6936fd232a7e4ee931255250eb08a4a16eff32635a1f8c1"
5
+ },
6
+ "WISP_en_captions.json": {
7
+ "bytes": 11864,
8
+ "sha256": "0f149bb3a8433feba1782b7a90517b14c2384e7048a1adf7282e52b414491409"
9
+ },
10
+ "WISP_MiMo_reproducible_project.zip": {
11
+ "bytes": 57091240,
12
+ "sha256": "797c20c808bf1db672ca298d55c4029630d9ccb16f8f8cf78e99869052349a96"
13
+ },
14
+ "WISP_en_mimo_voice.wav": {
15
+ "bytes": 21702444,
16
+ "sha256": "5707f14bbc83074c5aeea409c88faa1faed9ad6d7c804cd8857d7150c1e80081"
17
+ },
18
+ "WISP_en_cover.png": {
19
+ "bytes": 137541,
20
+ "sha256": "15d7c187ce4de39131b93cdf7b3bbea7b91cf350b2be7007d5a23d5d02ea2cfe"
21
+ },
22
+ "WISP_zh_script.txt": {
23
+ "bytes": 3300,
24
+ "sha256": "374bac0aa0158984875d3f454f6b8095dd0a3ae59d94aa250e3b56010495f2fe"
25
+ },
26
+ "WISP_zh_captions.json": {
27
+ "bytes": 9479,
28
+ "sha256": "ae495baff05ff44a11da7850fd721138f18c37836dd45454c416f2550750a03a"
29
+ },
30
+ "WISP_en_script.txt": {
31
+ "bytes": 3591,
32
+ "sha256": "70dc256ba3a5b61500090cdf7cc8dd80fac712c473c527cae3c10fc4c2bb51b7"
33
+ },
34
+ "WISP_zh_mimo_voice.wav": {
35
+ "bytes": 21804844,
36
+ "sha256": "3b9ef2e4d01bc0527ecc2ac2edfc5377c0a261b05e8d39e08ed9fd6d8e4f5726"
37
+ },
38
+ "WISP_en.vtt": {
39
+ "bytes": 4923,
40
+ "sha256": "51116e210153eed9fcbe026feb02083008c4d6fc6368f3beaf6fe51b84b38ba9"
41
+ },
42
+ "WISP_zh.srt": {
43
+ "bytes": 4340,
44
+ "sha256": "5a1bc0d4a390a592696e9a9868a6025915b81516a27ad0018b45955907243429"
45
+ },
46
+ "WISP_en.srt": {
47
+ "bytes": 5092,
48
+ "sha256": "fdc21df2bbc6cdbbae1252817ee48a36dc0c6afb1c504d9a7b2e60a2e285ed09"
49
+ },
50
+ "WISP_zh_cover.png": {
51
+ "bytes": 131082,
52
+ "sha256": "f7f33cdfcef368d2d9f484342dae5c7ab947dce60e1cf3f1890e63cab71c0775"
53
+ },
54
+ "WISP_en_mimo_voice.mp4": {
55
+ "bytes": 102486414,
56
+ "sha256": "6a84479dd6c90c65300f8a681c7dcf3a9cc2ef4cb023512f3ebf3924f3596cb1"
57
+ },
58
+ "WISP_zh.vtt": {
59
+ "bytes": 4216,
60
+ "sha256": "a1ba431ea7c28328dea30d38ae9003888be40577b8e0f730f3c1ef39a9379cb7"
61
+ }
62
+ }
videos/voice_release_20261028.json ADDED
@@ -0,0 +1,17 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "date": "2026-10-28",
3
+ "model": "mimo-v2.5-tts",
4
+ "voices": {
5
+ "zh": "茉莉",
6
+ "en": "Chloe"
7
+ },
8
+ "ai_generated_narration": true,
9
+ "seconds": {
10
+ "zh": 227.13333333333333,
11
+ "en": 226.06666666666666
12
+ },
13
+ "caption_alignment": "Unchanged source text aligned to local Whisper word timestamps; corrected English CIS subtitle splitting.",
14
+ "graphics_source_date": "2026-10-18",
15
+ "legacy_urls": "Existing system_voice.mp4 URLs are preserved as compatibility filenames for the refreshed editions.",
16
+ "old_release": "Preserved in local wisp_animation_20261018 and WISP/videos/archive/20261023."
17
+ }