@camstack/develop 0.1.6 → 0.1.8
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/package.json +1 -1
- package/skills/track-review.md +42 -8
package/package.json
CHANGED
package/skills/track-review.md
CHANGED
|
@@ -39,9 +39,20 @@ lacks it. Tokens are never printed.
|
|
|
39
39
|
|
|
40
40
|
## 1. Triage — rank, do not judge
|
|
41
41
|
|
|
42
|
-
`track_review_triage { hours, deviceIds?, classes? }` pages the
|
|
43
|
-
72 h, at most 14 days), scores every video track, and returns the
|
|
44
|
-
with their features
|
|
42
|
+
`track_review_triage { hours, deviceIds?, classes?, excludeInTraining? }` pages the
|
|
43
|
+
window (default 72 h, at most 14 days), scores every video track, and returns the
|
|
44
|
+
top suspects with their features and `cameraName`, plus a **random control sample**
|
|
45
|
+
of tracks that scored low.
|
|
46
|
+
|
|
47
|
+
Each row carries `inTraining: { collections } | null`, read from the retrain corpus
|
|
48
|
+
in one bounded pass (never per track). Suspects already in the corpus are skipped
|
|
49
|
+
by default and counted in `skippedInTraining`; pass `excludeInTraining: false` to
|
|
50
|
+
see them anyway. If `trainingTruncated` is true the corpus was only partly read, so
|
|
51
|
+
`null` is not proof.
|
|
52
|
+
|
|
53
|
+
`few-observations` is **not evaluated by triage**: the slim rows it reads carry no
|
|
54
|
+
position history and fetching it would cost a call per track. The result says so in
|
|
55
|
+
`signalsNotEvaluated`; `track_review_frame` shows the observation count for one track.
|
|
45
56
|
|
|
46
57
|
The score is a sum of cheap signals. None of them says "wrong"; together they say
|
|
47
58
|
"look here first":
|
|
@@ -76,6 +87,12 @@ cap: narrow it.
|
|
|
76
87
|
subject in the picture, not the label. If the frame is missing or too small to
|
|
77
88
|
tell, the verdict is `unsure`.
|
|
78
89
|
|
|
90
|
+
A whole-scene frame (key frame, full frame) is preferred, and the envelope is in
|
|
91
|
+
that image's coordinates. If only the track's thumbnail CROP fits the size bound,
|
|
92
|
+
`imageSpace` is `crop`, a `WARNING` says so, and the envelope is withheld: it is in
|
|
93
|
+
full-frame coordinates and the crop's rectangle is unknown, so it would not line up.
|
|
94
|
+
Raise `maxBytes` to get the full frame.
|
|
95
|
+
|
|
79
96
|
## 3. Verdicts
|
|
80
97
|
|
|
81
98
|
| Verdict | Meaning | Written? |
|
|
@@ -104,16 +121,17 @@ alongside the findings.
|
|
|
104
121
|
|
|
105
122
|
## 5. Plan, show, then apply
|
|
106
123
|
|
|
107
|
-
1. `track_review_plan { verdicts }` — local, writes nothing. It
|
|
108
|
-
|
|
124
|
+
1. `track_review_plan { verdicts, collectionName? }` — local, writes nothing. It
|
|
125
|
+
keeps each verdict's `note`, carries `collectionName` so the plan is
|
|
126
|
+
self-contained, and returns, per track, the steps (flag for training → copy the frame → save the corrected
|
|
109
127
|
boxes → mark complete → add to the collection), the subjects, and the
|
|
110
128
|
verdicts that write nothing with their reason.
|
|
111
129
|
2. **Present that write list to the operator and wait.** Name the tracks, the
|
|
112
130
|
classes you will write, the collection, and how many frames will be marked
|
|
113
131
|
complete. This is the only moment to catch a wrong call cheaply.
|
|
114
|
-
3. `track_review_apply { plan, collectionName
|
|
115
|
-
operator agrees. It refuses without `confirm: true` and
|
|
116
|
-
|
|
132
|
+
3. `track_review_apply { plan, collectionName?, confirm: true }` — only after the
|
|
133
|
+
operator agrees. It refuses without `confirm: true` and a collection name:
|
|
134
|
+
the argument if given, else `plan.collectionName` (the argument wins).
|
|
117
135
|
|
|
118
136
|
Apply is safe to repeat: a frame already copied is reused, a frame that already
|
|
119
137
|
has annotations is left alone (`overwrite: true` replaces them — a save replaces
|
|
@@ -122,6 +140,22 @@ fails does not stop the others, and an authorisation failure stops the run and
|
|
|
122
140
|
shows the hub's message. It returns a per-track log and the corpus KPI
|
|
123
141
|
(total / annotated / complete frames) before and after; quote both.
|
|
124
142
|
|
|
143
|
+
A box taken from the pipeline is saved as the pipeline's (`source: 'pipeline'`,
|
|
144
|
+
with its model, score and the `origin` it was offered as), so the hub can tell
|
|
145
|
+
"kept as drawn" from "relabelled". The save names the frame revision apply read;
|
|
146
|
+
if the operator edited the same frame in the admin page meanwhile, the hub
|
|
147
|
+
refuses it (`retrain-revision-conflict`) and that track is reported `failed` —
|
|
148
|
+
re-plan it rather than forcing it.
|
|
149
|
+
|
|
150
|
+
### Annotations saved before origins existed
|
|
151
|
+
|
|
152
|
+
`track_review_backfill_origin { collectionName | frameIds }` is a **dry-run**: per
|
|
153
|
+
frame it lists the rows that would get an origin (an EXACT rectangle match against
|
|
154
|
+
the pipeline box the hub re-derives), the rows that stay unmatched, and the rows
|
|
155
|
+
that already have one. Show that to the operator; only then repeat it with
|
|
156
|
+
`confirm: true`. It writes the origin column only and never overwrites an existing
|
|
157
|
+
origin. Rows left unmatched keep an approximate review verdict.
|
|
158
|
+
|
|
125
159
|
## Done when
|
|
126
160
|
|
|
127
161
|
- every suspect and every control track has a verdict;
|