labelled VIDEO TIME clock + time axis

app.py

@@ -28,11 +28,9 @@ DIMENSIONS = [
     ("localization", "2. Localization — do the red box / green mask cover the right object (not a neighbour, hand, or background)?"),
     ("state_history", "3. Location & state history — is the where / relation / timing (`now`, `history`) correct?"),
     ("moves", "4. Moves — are the recorded movement events (from → to, time) correct? (If `moves` is empty: is it correct that the object did not move?)"),
-    ("narration", "5. Narrations — do the blue NARRATION texts (HD-EPIC annotator descriptions, with their start/end times) really describe this object, and are the times right?"),
-    ("sound_speech", "6. Sound & speech — do the green SOUND events and any purple SPEECH (ASR of the audio) really belong to this object?"),
-    ("gaze", "7. Gaze — is the `looked at` claim right, i.e. does the red gaze dot sit on this object during the yellow `dwell` intervals?"),
+    ("sound_speech", "5. Sound & speech — do the green SOUND events and any purple SPEECH (ASR of the audio) really belong to this object?"),
+    ("gaze", "6. Gaze — is the `looked at` claim right, i.e. does the red gaze dot sit on this object during the yellow `dwell` intervals?"),
 ]
-COLUMNS = ["timestamp", "annotator", "video_id", "chunk", "object_id"] + [d[0] for d in DIMENSIONS] + ["comment"]
 
 # ---- optional persistence to a HF Dataset (used on HF Spaces, where local disk is ephemeral) ----
 DATASET_REPO = os.environ.get("RESPONSES_DATASET")   # e.g. WitneyWW/object-memory-eval-responses
@@ -53,6 +51,42 @@ if DATASET_REPO and os.environ.get("HF_TOKEN"):
                                 allow_patterns=["*.csv"])
 
 ITEMS = json.load(open(MEDIA / "index.json"))
+# Narrations are judged one by one. An object has between 0 and NARR_MAX of them, so the
+# page holds NARR_MAX radio groups and shows only as many as the current object needs.
+NARR_MAX = max(1, max(len(it["events"]["narration"]) for it in ITEMS))
+COLUMNS = (["timestamp", "annotator", "video_id", "chunk", "object_id"] + [d[0] for d in DIMENSIONS]
+           + ["narration_ids"] + [f"narration_{k + 1}" for k in range(NARR_MAX)] + ["comment"])
+NO_NARR_Q = ("No narration is linked to this object. Is that right — is there really no narration in this clip "
+             "about it? (All narrations of the clip are listed in the cross-check panel.)")
+
+# A responses file written with a different column set (older version of the form) is kept
+# under another name instead of being appended to with misaligned columns.
+if RESP_CSV.exists():
+    with open(RESP_CSV, newline="") as _f:
+        _hdr = next(csv.reader(_f), None)
+    if _hdr != COLUMNS:
+        RESP_CSV.rename(RESP_DIR / f"responses_old_{datetime.now():%Y%m%d_%H%M%S}.csv")
+
+
+def narrations_of(it):
+    return sorted(it["events"]["narration"], key=lambda e: e["t0"])
+
+
+def narration_updates(it):
+    """gr.update for each of the NARR_MAX radio groups: label + visibility for this object."""
+    ns = narrations_of(it)
+    ups = []
+    for k in range(NARR_MAX):
+        if k < len(ns):
+            e = ns[k]
+            ups.append(gr.update(visible=True, value=None,
+                                 label=f"Narration {k + 1} of {len(ns)} — [{e['t0']:.1f}–{e['t1']:.1f}s] “{e['label']}” — "
+                                       "is this about THIS object, and is the time right?"))
+        elif k == 0:
+            ups.append(gr.update(visible=True, value=None, label=NO_NARR_Q))
+        else:
+            ups.append(gr.update(visible=False, value=None))
+    return ups
 N = len(ITEMS)
 
 
@@ -169,7 +203,8 @@ def show(idx, annotator):
     done = load_done(annotator) if annotator else set()
     status = "✅ answered" if item_key(it) in done else "⬜ not answered"
     header = f"### {annotator or '?'} — item {idx + 1} / {N} &nbsp; {status} &nbsp; ({len(done)} / {N} done)"
-    return [idx, header, str(MEDIA / it["clip"]), gallery_for(it), annotation_md(it), crosscheck_md(it)] + [None] * len(DIMENSIONS) + [""]
+    return ([idx, header, str(MEDIA / it["clip"]), gallery_for(it), annotation_md(it), crosscheck_md(it)]
+            + [None] * len(DIMENSIONS) + narration_updates(it) + [""])
 
 
 def first_unanswered(annotator):
@@ -193,8 +228,9 @@ with gr.Blocks(title="Object-centric annotation eval") as demo:
     with gr.Column(visible=False) as main_col:
         header = gr.Markdown()
         gr.Markdown(
-            "All times are **seconds in the full recording** (shown top-right of the video and on the timeline), "
-            "not seconds since the clip started. "
+            "⏱ **All annotation times are seconds in the full recording.** The current time is the boxed "
+            "**VIDEO TIME** at the top-right of the video and the white playhead on the bottom time axis "
+            "(the player's own 0–30 s counter is NOT the annotation time). "
             "**Clip overlays:** red box = object near each keyframe · red dot = where the wearer is looking (gaze) · "
             "yellow *GAZE ON OBJECT* = the memory claims a gaze dwell · **blue NARRATION** = HD-EPIC annotator text "
             "(not audio) with its [start-end] and START/END markers · **purple SPEECH** = words actually spoken in the "
@@ -212,6 +248,9 @@ with gr.Blocks(title="Object-centric annotation eval") as demo:
             xc_md = gr.Markdown()
         gr.Markdown("## Your judgement")
         radios = [gr.Radio(CHOICES, label=q) for _, q in DIMENSIONS]
+        gr.Markdown("### Narrations — judge each one separately\n"
+                    "Blue NARRATION subtitles in the video; compare the [start–end] with the VIDEO TIME clock.")
+        narr_radios = [gr.Radio(CHOICES, label=f"Narration {k + 1}", visible=(k == 0)) for k in range(NARR_MAX)]
         comment = gr.Textbox(label="Comment (optional)", lines=2)
         with gr.Row():
             prev_btn = gr.Button("◀ Prev")
@@ -219,7 +258,7 @@ with gr.Blocks(title="Object-centric annotation eval") as demo:
             submit_btn = gr.Button("Submit & Next ▶", variant="primary")
         msg = gr.Markdown()
 
-    outputs = [idx_state, header, video, gallery, ann_md, xc_md] + radios + [comment]
+    outputs = [idx_state, header, video, gallery, ann_md, xc_md] + radios + narr_radios + [comment]
 
     def start(name):
         name = (name or "").strip()
@@ -235,20 +274,24 @@ with gr.Blocks(title="Object-centric annotation eval") as demo:
     skip_btn.click(lambda i, n: show(i + 1, n), [idx_state, name_state], outputs)
 
     def submit(i, name, *vals):
-        *answers, cmt = vals
-        if any(a is None for a in answers):
-            raise gr.Error(f"Please answer all {len(DIMENSIONS)} questions before submitting.")
+        answers, narr_ans, cmt = vals[:len(DIMENSIONS)], vals[len(DIMENSIONS):-1], vals[-1]
         it = ITEMS[i]
+        ns = narrations_of(it)
+        need = max(1, len(ns))
+        if any(a is None for a in answers) or any(a is None for a in narr_ans[:need]):
+            raise gr.Error(f"Please answer all {len(DIMENSIONS)} questions and all {need} narration question(s) before submitting.")
         row = {"timestamp": datetime.now().isoformat(timespec="seconds"), "annotator": name,
                "video_id": it["video_id"], "chunk": it["chunk"], "object_id": it["object_id"],
                "comment": (cmt or "").strip()}
         row.update({d[0]: a for d, a in zip(DIMENSIONS, answers)})
+        row["narration_ids"] = ";".join(str(e.get("id")) for e in ns) or "(none)"
+        row.update({f"narration_{k + 1}": (narr_ans[k] if k < need else "") for k in range(NARR_MAX)})
         append_row(row)
         if i + 1 >= N:
             return show(i, name) + [f"**Saved.** That was the last item — all {N} done. Thank you!"]
         return show(i + 1, name) + [f"Saved `{it['object_id']}` → {RESP_CSV.name}"]
 
-    submit_btn.click(submit, [idx_state, name_state] + radios + [comment], outputs + [msg])
+    submit_btn.click(submit, [idx_state, name_state] + radios + narr_radios + [comment], outputs + [msg])
 
 if __name__ == "__main__":
     ap = argparse.ArgumentParser()