Replace pick_barrier.py with thumbnail-grid UX

Old version showed inter-fly distance plots and asked the analyst to
click a timeline. The new version reads frames directly from the .mp4
and shows a 10×6 grid of timestamped thumbnails — the analyst just
clicks the frame where the barrier opens.

Two-stage refinement:
  - Coarse grid: 60 thumbs spanning the 5-min search window at ~5 s
    spacing. Pick the rough moment.
  - Fine grid: 60 thumbs at 0.2 s spacing centred on the coarse pick.
    Pick the exact frame.

Auto-detector still feeds the starting position. Sequential video
decode (one cv2 pass through the relevant range) instead of seek-per-
frame, so each grid loads in a few seconds.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

scripts/pick_barrier.py

@@ -1,47 +1,48 @@
-"""Interactive picker for barrier-opening time per tracked video.
+"""Interactive picker for barrier-opening time, frame-by-frame thumbnail style.
 
-Loops through tracked DBs that don't yet have a barrier-opening
+For each video that doesn't yet have a barrier-opening annotation, show a
-annotation. For each, plots the windowed mean inter-fly distance for
+10x6 grid of timestamped thumbnails extracted directly from the .mp4.
-all 6 ROIs over the first 5 minutes and lets the analyst click the
+The analyst clicks the thumbnail at (or just after) the moment the
-moment the barrier opens (when most flies start coming close together).
+barrier opens; the picker then refines with a second tighter grid for
-The auto-detector's best-effort guess is shown as a starting position.
+sub-second precision.
+
+Two-stage flow per video:
+    1. Coarse grid: 60 thumbs spanning the 5-min search window (5 s spacing).
+       Click → pick that 5 s slot.
+    2. Fine grid:   60 thumbs spanning ±6 s of the coarse pick (0.2 s spacing).
+       Click → final answer with 0.2 s precision.
 
 Output: data/metadata/barrier_opening.csv with columns
     machine_name, session_date, session_time, opening_s, trim_first_s, notes
 
-`opening_s` is the moment of barrier opening, measured from the start
-of the recording (NOT the start of any trimmed copy). `trim_first_s`
-is an optional annotation for videos with a misframed start that
-should be ignored by analysis (defaults to 0).
-
 Window keys:
-    click        place the opening cursor at that time
+    click        select thumbnail at that timestamp
-    ENTER        save and advance
+    n            skip this video for THIS run
-    [, ]         shift cursor by 1 s left / right
+    u            mark unusable (opening_s = NaN)
-    {, }         shift cursor by 5 s left / right
+    b            back to coarse grid (after seeing fine grid)
-    n            skip this video for THIS run (no row written)
-    u            mark this video unusable (writes opening_s = NaN, notes = "unusable")
-    r            reset cursor to the auto-detected position
     q / ESC      save+quit
 
 Usage:
     python pick_barrier.py
-    python pick_barrier.py --redo   # re-pick videos that already have a row
+    python pick_barrier.py --redo
     python pick_barrier.py --limit 10
+    python pick_barrier.py --db /path/to/specific_tracking.db
 """
 
 from __future__ import annotations
 
 import argparse
+import re
 import sqlite3
 import sys
 from pathlib import Path
 
+import cv2
 import matplotlib.pyplot as plt
 import numpy as np
 import pandas as pd
 
-from config import DATA_METADATA, VIDEO_INFO_TSV
+from config import DATA_METADATA, INVENTORY_CSV, VIDEO_INFO_TSV
 from detect_barrier_opening import (
     SEARCH_END_S, STEP_S, WINDOW_S,
     per_frame_distance, sliding_mean,
@@ -51,43 +52,30 @@ OUT_CSV = DATA_METADATA / "barrier_opening.csv"
 OUT_COLS = ["machine_name", "session_date", "session_time",
             "opening_s", "trim_first_s", "notes"]
 
-
+DB_NAME_RE = re.compile(
-def parse_db_filename(db_path: Path) -> tuple[str, str, str] | None:
+    r"^(\d{4}-\d{2}-\d{2})_(\d{2}-\d{2}-\d{2})_([0-9a-f]{32})__"
-    """Pull (date, time, machine_uuid) out of a tracking DB filename."""
-    import re
-    m = re.match(
-        r"^(\d{4}-\d{2}-\d{2})_(\d{2}-\d{2}-\d{2})_([0-9a-f]{32})__",
-        db_path.name,
 )
-    if not m:
+
+GRID_ROWS, GRID_COLS = 6, 10
+N_THUMBS = GRID_ROWS * GRID_COLS  # 60
+COARSE_SPAN_S = SEARCH_END_S       # 0..300s, ~5s spacing
+FINE_SPAN_S = 12.0                 # ±6s around coarse pick → ~0.2s spacing
+
+
+def auto_suggest(db_path: Path) -> float | None:
+    """Median of per-ROI biggest-drop times. None if too noisy."""
+    try:
+        conn = sqlite3.connect(f"file:{db_path}?mode=ro", uri=True)
+    except sqlite3.Error:
         return None
-    return m.group(1), m.group(2), m.group(3)
+    candidates = []
-
-
-def load_distance_traces(db_path: Path) -> dict[int, pd.DataFrame]:
-    """For each ROI 1..6, return windowed-mean DF; empty if ROI missing."""
-    out: dict[int, pd.DataFrame] = {}
-    with sqlite3.connect(f"file:{db_path}?mode=ro", uri=True) as conn:
     for roi in range(1, 7):
         try:
-                df = pd.read_sql_query(
+            df = pd.read_sql_query(f"SELECT t, x, y, id FROM ROI_{roi}", conn)
-                    f"SELECT t, x, y, id FROM ROI_{roi}", conn
-                )
         except Exception:
-                out[roi] = pd.DataFrame()
             continue
         dist = per_frame_distance(df)
-            out[roi] = sliding_mean(dist, WINDOW_S, STEP_S, SEARCH_END_S)
+        smean = sliding_mean(dist, WINDOW_S, STEP_S, SEARCH_END_S)
-    return out
-
-
-def auto_suggest(traces: dict[int, pd.DataFrame]) -> float | None:
-    """Median of per-ROI biggest-drop times. Returns None if too noisy."""
-    candidates = []
-    for roi, smean in traces.items():
-        if len(smean) < 30:
-            continue
-        # Find the time of the largest decrease in median(pre)–median(post).
         pad = max(1, int(WINDOW_S / STEP_S))
         if len(smean) < 2 * pad + 1:
             continue
@@ -102,103 +90,200 @@ def auto_suggest(traces: dict[int, pd.DataFrame]) -> float | None:
                 best_t = float(smean["mid_t"].iloc[i])
         if best_drop > 30 and best_t is not None:
             candidates.append(best_t)
+    conn.close()
     if not candidates:
         return None
     return float(np.median(candidates))
 
 
-def show_picker(
+def grab_thumbnails(video_path: Path, target_times_s: np.ndarray,
-    db_path: Path,
+                    thumb_w: int = 320) -> list[np.ndarray | None]:
-    machine_name: str,
+    """Read thumbnails at the requested timestamps via a single sequential pass.
-    session_date: str,
-    session_time: str,
-    auto_t: float | None,
-    initial_t: float,
-) -> dict | None:
-    """Open the picker window. Returns a dict ready for OUT_CSV, or None to skip."""
-    traces = load_distance_traces(db_path)
-    if all(s.empty for s in traces.values()):
-        print(f"  ! no usable ROI traces in {db_path.name}; skipping")
-        return None
 
-    fig, axes = plt.subplots(6, 1, figsize=(13, 12), sharex=True)
+    Linear-decode is much faster than seeking per-frame on H.264. We read
-    fig.suptitle(
+    frames sequentially from the earliest target onward, keeping only the
-        f"{machine_name}  {session_date} {session_time}\n"
+    ones at requested target frames.
-        f"click ↦ set opening · ENTER save · "
+    """
-        f"[/] ±1s · {{/}} ±5s · n skip · u unusable · r reset · q quit",
+    cap = cv2.VideoCapture(str(video_path))
-        fontsize=10,
+    fps = cap.get(cv2.CAP_PROP_FPS) or 25.0
+    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+    src_w = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+    src_h = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+    if total_frames <= 0:
+        cap.release()
+        return [None] * len(target_times_s)
+
+    target_frames = np.clip(
+        (target_times_s * fps).round().astype(int), 0, total_frames - 1
     )
+    sort_idx = np.argsort(target_frames)
+    sorted_targets = target_frames[sort_idx]
 
-    state: dict = {"t": float(initial_t), "auto": auto_t, "result": None}
+    out: list[np.ndarray | None] = [None] * len(target_times_s)
+    if sorted_targets.size == 0:
+        cap.release()
+        return out
 
-    def redraw():
+    cap.set(cv2.CAP_PROP_POS_FRAMES, int(sorted_targets[0]))
-        for ax, (roi, smean) in zip(axes, sorted(traces.items())):
+    cur_frame = int(sorted_targets[0])
-            ax.cla()
+    last_frame_data: np.ndarray | None = None
-            if smean.empty:
+
-                ax.text(0.5, 0.5, f"ROI {roi}: no data",
+    scale = thumb_w / src_w if src_w > 0 else 1.0
-                        transform=ax.transAxes, ha="center", va="center", color="grey")
+    thumb_h = max(1, int(round(src_h * scale)))
+
+    for ord_i, target in zip(sort_idx, sorted_targets):
+        while cur_frame <= target:
+            ret, frame = cap.read()
+            if not ret:
+                last_frame_data = None
+                break
+            last_frame_data = frame
+            cur_frame += 1
+        if last_frame_data is not None:
+            small = cv2.resize(last_frame_data, (thumb_w, thumb_h),
+                               interpolation=cv2.INTER_AREA)
+            out[ord_i] = cv2.cvtColor(small, cv2.COLOR_BGR2RGB)
+
+    cap.release()
+    return out
+
+
+def show_thumbnail_grid(
+    video_path: Path,
+    center_t: float,
+    span_s: float,
+    title: str,
+) -> tuple[float | None, str]:
+    """Show a 10×6 thumbnail grid; return (clicked_time, action).
+
+    `action` is one of: 'pick', 'skip', 'unusable', 'back', 'quit'.
+    `clicked_time` is None unless action == 'pick'.
+    """
+    half = span_s / 2.0
+    times = np.linspace(max(0.0, center_t - half), center_t + half, N_THUMBS)
+    print(f"  loading {N_THUMBS} thumbnails ({times[0]:.1f}–{times[-1]:.1f}s)...", flush=True)
+    thumbs = grab_thumbnails(video_path, times)
+
+    fig, axes = plt.subplots(GRID_ROWS, GRID_COLS, figsize=(20, 11))
+    fig.suptitle(
+        f"{title}\nclick a thumbnail · n=skip · u=unusable · b=back · q=quit",
+        fontsize=11,
+    )
+    state = {"time": None, "action": None}
+
+    for ax, t, thumb in zip(axes.flat, times, thumbs):
+        if thumb is not None:
+            ax.imshow(thumb)
         else:
-                ax.plot(smean["mid_t"], smean["mean_dist"], color="steelblue", lw=1.0)
+            ax.set_facecolor("black")
-                ax.set_ylabel(f"ROI {roi}")
+            ax.text(0.5, 0.5, "no frame",
-            if state["auto"] is not None:
+                    transform=ax.transAxes, ha="center", va="center", color="white")
-                ax.axvline(state["auto"], color="orange", ls=":", lw=0.8, alpha=0.8)
+        # Format time as M:SS.s for readability
-            ax.axvline(state["t"], color="red", lw=1.5)
+        m, s = divmod(t, 60)
-            ax.set_xlim(0, SEARCH_END_S)
+        ax.set_title(f"{int(m):d}:{s:05.2f}", fontsize=8, pad=1)
-            ax.grid(True, alpha=0.3)
+        ax.set_xticks([]); ax.set_yticks([])
-        axes[-1].set_xlabel("time (s)")
+    fig.subplots_adjust(left=0.01, right=0.99, top=0.93, bottom=0.01,
-        axes[0].set_title(f"orange dotted = auto-suggested  ·  red = current pick: {state['t']:.1f} s",
+                        wspace=0.03, hspace=0.18)
-                          fontsize=9)
-        fig.canvas.draw_idle()
 
     def on_click(event):
-        if event.inaxes in axes and event.xdata is not None:
+        if event.inaxes is None:
-            state["t"] = max(0.0, min(SEARCH_END_S, float(event.xdata)))
+            return
-            redraw()
+        for i, ax in enumerate(axes.flat):
+            if ax is event.inaxes:
+                state["time"] = float(times[i])
+                state["action"] = "pick"
+                plt.close(fig)
+                return
 
     def on_key(event):
         k = event.key
-        if k == "enter":
+        if k == "n":
-            state["result"] = {
+            state["action"] = "skip"; plt.close(fig)
-                "machine_name": machine_name,
-                "session_date": session_date,
-                "session_time": session_time,
-                "opening_s": round(state["t"], 1),
-                "trim_first_s": 0,
-                "notes": "",
-            }
-            plt.close(fig)
-        elif k == "n":
-            state["result"] = "skip"
-            plt.close(fig)
         elif k == "u":
-            state["result"] = {
+            state["action"] = "unusable"; plt.close(fig)
-                "machine_name": machine_name,
+        elif k == "b":
-                "session_date": session_date,
+            state["action"] = "back"; plt.close(fig)
-                "session_time": session_time,
-                "opening_s": np.nan,
-                "trim_first_s": 0,
-                "notes": "unusable",
-            }
-            plt.close(fig)
-        elif k == "r" and state["auto"] is not None:
-            state["t"] = state["auto"]
-            redraw()
-        elif k in ("[", "]"):
-            state["t"] = max(0.0, min(SEARCH_END_S, state["t"] + (-1 if k == "[" else 1)))
-            redraw()
-        elif k in ("{", "}"):
-            state["t"] = max(0.0, min(SEARCH_END_S, state["t"] + (-5 if k == "{" else 5)))
-            redraw()
         elif k in ("q", "escape"):
-            state["result"] = "quit"
+            state["action"] = "quit"; plt.close(fig)
-            plt.close(fig)
 
     fig.canvas.mpl_connect("button_press_event", on_click)
     fig.canvas.mpl_connect("key_press_event", on_key)
-    redraw()
     plt.show()
+    return state["time"], state["action"] or "skip"
 
-    return state["result"]
+
+def pick_for_video(
+    video_path: Path,
+    db_path: Path | None,
+    machine_name: str,
+    session_date: str,
+    session_time: str,
+) -> dict | str | None:
+    """Run the two-stage thumbnail picker. Return dict, 'skip', or 'quit'."""
+    auto_t = auto_suggest(db_path) if db_path else None
+    print(f"  auto-suggest: {f'{auto_t:.1f}s' if auto_t else '(none)'}")
+
+    # Stage 1: coarse grid centred on auto-suggest (or 150 s default).
+    coarse_center = auto_t if auto_t is not None else COARSE_SPAN_S / 2
+    title_coarse = f"COARSE  {machine_name}  {session_date} {session_time}  ·  spanning 5 min"
+    while True:
+        coarse_t, action = show_thumbnail_grid(
+            video_path, coarse_center, COARSE_SPAN_S, title_coarse
+        )
+        if action == "skip":
+            return "skip"
+        if action == "unusable":
+            return {
+                "machine_name": machine_name, "session_date": session_date,
+                "session_time": session_time, "opening_s": np.nan,
+                "trim_first_s": 0, "notes": "unusable",
+            }
+        if action == "quit":
+            return "quit"
+        if action == "back":
+            continue   # already at the top stage; redraw
+        if action == "pick" and coarse_t is not None:
+            break
+
+    # Stage 2: fine grid around the coarse pick.
+    title_fine = (f"FINE  {machine_name}  {session_date} {session_time}  "
+                  f"·  ±{FINE_SPAN_S/2:.0f} s around {coarse_t:.1f} s")
+    while True:
+        fine_t, action = show_thumbnail_grid(
+            video_path, coarse_t, FINE_SPAN_S, title_fine
+        )
+        if action == "back":
+            return pick_for_video(video_path, db_path, machine_name,
+                                  session_date, session_time)
+        if action == "skip":
+            return "skip"
+        if action == "unusable":
+            return {
+                "machine_name": machine_name, "session_date": session_date,
+                "session_time": session_time, "opening_s": np.nan,
+                "trim_first_s": 0, "notes": "unusable",
+            }
+        if action == "quit":
+            return "quit"
+        if action == "pick" and fine_t is not None:
+            return {
+                "machine_name": machine_name, "session_date": session_date,
+                "session_time": session_time, "opening_s": round(fine_t, 1),
+                "trim_first_s": 0, "notes": "",
+            }
+
+
+def lookup_video_path(machine_name: str, session_date: str,
+                      session_time: str, inv: pd.DataFrame) -> Path | None:
+    """Find the mp4 path for (machine, date, time) in the inventory."""
+    match = inv[
+        (inv["machine_name"] == machine_name)
+        & (inv["session_date"] == session_date)
+        & (inv["session_time"] == session_time)
+    ]
+    if match.empty:
+        return None
+    return Path(match.iloc[0]["mp4_path"])
 
 
 def main() -> None:
@@ -208,7 +293,7 @@ def main() -> None:
     parser.add_argument("--limit", type=int, default=None,
                         help="only process the first N videos")
     parser.add_argument("--db", type=Path, default=None,
-                        help="annotate this specific DB only")
+                        help="annotate this specific tracking DB only")
     args = parser.parse_args()
 
     OUT_CSV.parent.mkdir(parents=True, exist_ok=True)
@@ -218,10 +303,15 @@ def main() -> None:
         out = pd.DataFrame(columns=OUT_COLS)
     done = set(zip(out["machine_name"], out["session_date"], out["session_time"]))
 
-    # Build the queue: every tracked DB referenced by the merged TSV that
+    if not INVENTORY_CSV.exists():
-    # hasn't been picked yet.
+        sys.exit(f"Inventory not found at {INVENTORY_CSV}. Run build_video_inventory.py first.")
+    inv = pd.read_csv(INVENTORY_CSV)
+
+    # Build the queue: every (machine, date, time) referenced by the merged
+    # TSV that has a tracking DB on disk and isn't yet annotated.
     tsv = pd.read_csv(VIDEO_INFO_TSV, sep="\t")
-    queue = []
+    queue: list[tuple[Path, Path, str, str, str]] = []
+    seen: set[tuple[str, str, str]] = set()
     for col in ("training_db_path", "testing_db_path"):
         for _, row in tsv.iterrows():
             db = row[col]
@@ -230,28 +320,25 @@ def main() -> None:
             db_path = Path(db)
             if not db_path.exists():
                 continue
-            parsed = parse_db_filename(db_path)
+            m = DB_NAME_RE.match(db_path.name)
-            if parsed is None:
+            if not m:
                 continue
-            session_date, session_time, _ = parsed
+            session_date, session_time = m.group(1), m.group(2)
             key = (row["machine_name"], session_date, session_time)
+            if key in seen:
+                continue
+            seen.add(key)
             if key in done and not args.redo:
                 continue
-            queue.append((db_path, row["machine_name"], session_date, session_time))
+            video = lookup_video_path(*key, inv)
-
+            if video is None or not video.exists():
-    # Dedup (a fly may reference the same DB for both training & testing).
+                print(f"  ! no video for {key}; skipping")
-    seen = set()
+                continue
-    deduped = []
+            queue.append((db_path, video, *key))
-    for item in queue:
-        k = (item[1], item[2], item[3])
-        if k not in seen:
-            seen.add(k)
-            deduped.append(item)
-    queue = deduped
 
     if args.db:
         target = Path(args.db).resolve()
-        queue = [q for q in queue if Path(q[0]).resolve() == target]
+        queue = [q for q in queue if q[0].resolve() == target]
         if not queue:
             sys.exit(f"DB not found in queue: {args.db}")
 
@@ -259,34 +346,26 @@ def main() -> None:
         queue = queue[: args.limit]
 
     if not queue:
-        print("Nothing to pick. All eligible DBs already have a barrier_opening row.")
+        print("Nothing to pick. All eligible videos already have a barrier_opening row.")
         return
 
     print(f"Picking barrier-opening for {len(queue)} videos.")
-    print("Window keys: click=set  ENTER=save  [/]=±1s  {/}=±5s  n=skip  u=unusable  r=reset  q=quit")
+    print("Window keys: click=pick · n=skip · u=unusable · b=back · q=quit")
 
     saved = skipped = unusable = 0
-    for i, (db, machine_name, session_date, session_time) in enumerate(queue, 1):
+    for i, (db, video, machine_name, session_date, session_time) in enumerate(queue, 1):
         prefix = f"[{i}/{len(queue)}] {machine_name} {session_date} {session_time}"
         print(f"\n{prefix}")
 
-        traces = load_distance_traces(db)
+        result = pick_for_video(video, db, machine_name, session_date, session_time)
-        auto_t = auto_suggest(traces)
-        initial = auto_t if auto_t is not None else 60.0
-        print(f"  auto-suggest: "
-              f"{f'{auto_t:.1f}s' if auto_t is not None else '(none)'}")
-
-        result = show_picker(db, machine_name, session_date, session_time,
-                             auto_t, initial)
 
         if result is None or result == "skip":
             skipped += 1
             continue
         if result == "quit":
-            print("  quit requested — saving what we have and exiting")
+            print("  quit requested — saving and exiting")
             break
 
-        # Append + dedup on key + persist after each save (crash-safe).
         new_row = pd.DataFrame([result])
         out = pd.concat([
             out[~((out.machine_name == result["machine_name"]) &
@@ -297,6 +376,7 @@ def main() -> None:
         out[OUT_COLS].to_csv(OUT_CSV, index=False)
         if pd.isna(result["opening_s"]):
             unusable += 1
+            print("  saved as unusable")
         else:
             saved += 1
             print(f"  saved opening_s = {result['opening_s']} s")