The live picture was chained to the recognition pipeline

Reported from the first Windows install: the camera feed lags. It did,
and not because of the network, the proxy or the webview.

The MJPEG stream served _annotated_jpeg - the frame the pipeline had
most recently FINISHED with, encoded after detection, quality scoring,
tracking and identification had all run on it. On a modest shop PC that
is a few frames a second, and every picture was already as old as that
processing. It looked like lag because it was lag. On the fast machine
it was developed on the pipeline kept up with the stream's own 10 fps
cap, which is why nobody here ever saw it.

Two more things compounded it. Every processed frame was JPEG-encoded
whether or not a viewer existed - CPU spent on precisely the machine
short of it. And ffmpeg ran its RTSP demuxer with default buffering,
which holds a comfortable queue of frames before handing over the first:
half a second to two seconds a live view can never recover.

Now the picture and the boxes are decoupled. latest_jpeg_since takes the
capture thread's freshest frame at the camera's own rate and draws the
boxes from the last processed frame over it - encoded on demand, per
request, so a camera nobody watches costs no encode at all. The stream
sends a frame only when the camera has a newer one, capped at 15 fps;
nothing is sent twice. Boxes older than a second are not drawn, so a
stalled pipeline cannot leave one floating over an empty spot.
_publish_annotated becomes _remember_tracks: a handful of tuples under
the lock, no copy, no encode. ffmpeg gets nobuffer / low_delay /
max_delay.

Measured on cam2's sub-stream, same machine, ten seconds each:

  before   99 frames sent,  98 distinct    9.8 new pictures/s
  after   141 frames sent, 141 distinct   14.0 new pictures/s

against a 15 fps camera, with the pipeline still processing 166 of 181
captured frames alongside - and engine CPU DOWN from 90% with no viewer
to 62% with one attached.

Engine version 1.0.0 -> 1.1.0 so a re-run of setup reinstalls it rather
than pip deciding the requirement is already satisfied.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01KGcjxF1cNLcuwc3DAPcnfj

behavision/api.py

@@ -6,6 +6,7 @@ models finish loading without a single unguarded None dereference.
 from __future__ import annotations
 
 import asyncio
+import time
 import logging
 import secrets
 from pathlib import Path
@@ -370,11 +371,23 @@ def create_app(engine: Engine) -> FastAPI:
             # Stop when the camera is deleted or its worker dies - otherwise a
             # removed camera leaves this generator running for the life of the
             # process, holding a reference to a worker nothing else can see.
+            # Driven by the camera, not a timer: a frame goes out when the
+            # capture thread has one newer than the last one sent, so nothing
+            # is sent twice and nothing waits on the recognition pipeline.
+            # Capped at 15 fps - the office cameras' own rate - so a viewer
+            # never costs more encodes than the camera produces pictures.
+            last_ts, min_gap, sent_at = 0.0, 1.0 / 15, 0.0
             while engine.workers.get(camera_id) is worker and worker.is_alive():
-                jpeg = worker.latest_jpeg()
-                if jpeg is not None:
-                    yield boundary + jpeg + b"\r\n"
-                await asyncio.sleep(0.1)  # ~10 fps to the browser
+                now = time.time()
+                if now - sent_at < min_gap:
+                    await asyncio.sleep(min_gap - (now - sent_at))
+                    continue
+                jpeg, ts = worker.latest_jpeg_since(last_ts)
+                if jpeg is None:
+                    await asyncio.sleep(0.02)
+                    continue
+                last_ts, sent_at = ts, time.time()
+                yield boundary + jpeg + b"\r\n"
 
         return StreamingResponse(
             generate(),

behavision/capture.py

@@ -17,10 +17,20 @@ import numpy as np
 
 log = logging.getLogger(__name__)
 
-# Force TCP transport and a 5s socket timeout for RTSP before OpenCV loads
-# ffmpeg. UDP is the default and silently drops frames on lossy Wi-Fi.
+# Set before OpenCV loads ffmpeg, which reads this once.
+#
+# rtsp_transport=tcp: UDP is the default and silently drops frames on lossy
+# Wi-Fi. stimeout: a 5s socket timeout so a dead camera is noticed.
+#
+# fflags=nobuffer and flags=low_delay: without them ffmpeg's RTSP demuxer
+# holds a comfortable queue of frames before handing over the first, which
+# on a live feed is half a second to two seconds of latency that no amount of
+# work downstream can recover - the frame is already old when we get it. A
+# recorder wants that buffer; a live view does not. max_delay caps the
+# reorder wait for the same reason.
 os.environ.setdefault(
-    "OPENCV_FFMPEG_CAPTURE_OPTIONS", "rtsp_transport;tcp|stimeout;5000000"
+    "OPENCV_FFMPEG_CAPTURE_OPTIONS",
+    "rtsp_transport;tcp|stimeout;5000000|fflags;nobuffer|flags;low_delay|max_delay;200000",
 )
 
 

behavision/engine.py

@@ -162,7 +162,11 @@ class CameraWorker(threading.Thread):
         # every test using a stubbed worker passed.
         self._stopping = threading.Event()
         self._lock = threading.Lock()
-        self._annotated_jpeg: Optional[bytes] = None
+        # What the live view draws over the freshest frame: the boxes from
+        # the most recent processed frame, and when they were computed. NOT a
+        # pre-rendered JPEG - see latest_jpeg for why.
+        self._overlay: "list[tuple[tuple[int, int, int, int], tuple[int, int, int], str]]" = []
+        self._overlay_ts = 0.0
         self._last_frame_ts = 0.0
         self._was_connected = False
         self.frames_processed = 0
@@ -187,8 +191,46 @@ class CameraWorker(threading.Thread):
         self.source.stop()
 
     def latest_jpeg(self) -> Optional[bytes]:
+        jpeg, _ = self.latest_jpeg_since(0.0)
+        return jpeg
+
+    def latest_jpeg_since(self, known_ts: float) -> "tuple[Optional[bytes], float]":
+        """The freshest captured frame with the latest boxes drawn on it, or
+        (None, known_ts) if the camera has produced nothing newer.
+
+        The live picture is deliberately NOT the frame the pipeline last
+        finished with. That version advanced only when detection, tracking and
+        identification had all completed on a frame - a few times a second on a
+        modest shop PC - and every picture it showed was already as old as that
+        processing. It looked like lag because it was lag. Here the picture runs
+        at the camera's rate off the capture thread's latest frame, and the
+        boxes - which genuinely can only update at pipeline rate - are drawn
+        over it from the last processed frame. Boxes may trail a fast walker by
+        one pipeline period; the picture never does.
+
+        Encoded on demand, per request, so a camera nobody is watching pays for
+        no JPEG at all. The old path encoded every processed frame whether or
+        not a viewer existed - CPU spent on precisely the machine short of it.
+        """
+        frame, ts = self.source.latest_since(known_ts)
+        if frame is None:
+            return None, known_ts
         with self._lock:
-            return self._annotated_jpeg
+            overlay, overlay_ts = list(self._overlay), self._overlay_ts
+        # A stalled pipeline must not leave a box floating over an empty spot.
+        # Older than a second and the person has walked out from under it.
+        draw = overlay if (time.time() - overlay_ts) < 1.0 else []
+        if draw:
+            frame = frame.copy()
+            for (x1, y1, x2, y2), color, text in draw:
+                cv2.rectangle(frame, (x1, y1), (x2, y2), color, 2)
+                if text:
+                    cv2.putText(frame, text, (x1, max(20, y1 - 8)),
+                                cv2.FONT_HERSHEY_SIMPLEX, 0.55, color, 2)
+        ok, buf = cv2.imencode(".jpg", frame, [int(cv2.IMWRITE_JPEG_QUALITY), 80])
+        if not ok:
+            return None, known_ts
+        return buf.tobytes(), ts
 
     def stats(self) -> dict:
         return {
@@ -236,7 +278,7 @@ class CameraWorker(threading.Thread):
                 for track in ended:
                     self._finish_track(track, ts)
 
-                self._publish_annotated(frame, active)
+                self._remember_tracks(active)
                 self.frames_processed += 1
             except Exception:
                 log.exception("[%s] frame processing failed", self.cam_cfg.id)
@@ -426,12 +468,14 @@ class CameraWorker(threading.Thread):
                                            track.quality, rcfg=self.rcfg):
             track.reinforcements += 1
 
-    def _publish_annotated(self, frame: np.ndarray, tracks: "list[Track]") -> None:
-        canvas = frame.copy()
+    def _remember_tracks(self, tracks: "list[Track]") -> None:
+        """Record what to draw. Cheap: a handful of tuples under the lock,
+        no frame copy and no encode. The encode happens in latest_jpeg_since,
+        only when somebody is looking."""
+        overlay = []
         for t in tracks:
             if t.misses > 0:
                 continue  # only draw tracks matched in this frame
-            x1, y1, x2, y2 = t.box
             if t.state == "resolved":
                 color = _COLORS["known"] if t.label and not str(t.label).startswith(
                     "Visitor") else _COLORS["new"]
@@ -440,15 +484,10 @@ class CameraWorker(threading.Thread):
                 color, text = _COLORS["ambiguous"], "?"
             else:
                 color, text = _COLORS["pending"], ""
-            cv2.rectangle(canvas, (x1, y1), (x2, y2), color, 2)
-            if text:
-                cv2.putText(canvas, text, (x1, max(20, y1 - 8)),
-                            cv2.FONT_HERSHEY_SIMPLEX, 0.55, color, 2)
-        ok, buf = cv2.imencode(".jpg", canvas,
-                               [int(cv2.IMWRITE_JPEG_QUALITY), 80])
-        if ok:
-            with self._lock:
-                self._annotated_jpeg = buf.tobytes()
+            overlay.append((tuple(t.box), color, text))
+        with self._lock:
+            self._overlay = overlay
+            self._overlay_ts = time.time()
 
 
 class Engine:

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "behavision"
-version = "1.0.0"
+version = "1.1.0"
 description = "Production face recognition over RTSP"
 requires-python = ">=3.10"
 dependencies = [

tests/test_api_cameras.py

@@ -31,6 +31,7 @@ class FakeWorker:
     def stats(self): return {"camera_id": self.cam_cfg.id, "connected": True,
                              "url": self.cam_cfg.safe_url()}
     def latest_jpeg(self): return None
+    def latest_jpeg_since(self, known_ts): return None, known_ts
 
 
 @pytest.fixture

tests/test_live_picture.py

@@ -0,0 +1,144 @@
+"""The live picture is the camera's latest frame, not the pipeline's.
+
+Until this, the MJPEG stream served the frame the recognition pipeline had
+most recently *finished* - so on a shop PC where detection plus identification
+ran a few times a second, the live view ran a few times a second too, and every
+picture it showed was already as old as that processing. It read as lag
+because it was. These tests pin the decoupling: the picture comes from the
+capture thread at its own rate, the boxes come from the pipeline at theirs,
+and nothing is encoded for a camera nobody is watching.
+"""
+import time
+
+import numpy as np
+
+from behavision.config import CameraConfig, Config
+from behavision.engine import CameraWorker
+from behavision.tracking import Track
+
+
+class StillSource:
+    """A capture thread stand-in that hands out whatever frame it is given."""
+
+    def __init__(self):
+        self.frame = None
+        self.ts = 0.0
+
+    def set(self, frame):
+        self.frame, self.ts = frame, time.time()
+
+    def latest(self):
+        return self.frame, self.ts
+
+    def latest_since(self, known_ts):
+        if self.frame is None or self.ts <= known_ts:
+            return None, known_ts
+        return self.frame, self.ts
+
+    def stop(self):
+        pass
+
+
+def make_worker(tmp_path):
+    cfg = Config()
+    cfg.app.data_dir = tmp_path
+    cam = CameraConfig(id="cam1", host="127.0.0.1", port=1, path="/none")
+    w = CameraWorker(cam, cfg, detector=None, encoder=None, gallery=None,
+                     bus=None, attrs=None)
+    w.source = StillSource()
+    return w
+
+
+def grey(v=90):
+    return np.full((120, 160, 3), v, dtype=np.uint8)
+
+
+def resolved_track(box=(30, 30, 90, 100), label="Priya"):
+    t = Track(id=1, box=box, kps=np.zeros((5, 2), dtype=np.float32), score=0.9)
+    t.state, t.label, t.similarity = "resolved", label, 0.71
+    return t
+
+
+def test_the_picture_advances_with_the_camera_not_the_pipeline(tmp_path):
+    w = make_worker(tmp_path)
+
+    # Nothing captured yet: nothing to show, and no encode happened.
+    assert w.latest_jpeg_since(0.0) == (None, 0.0)
+
+    # A frame arrives from the camera. The pipeline has not touched it - and
+    # the live view must not wait for it to.
+    w.source.set(grey(80))
+    jpeg1, ts1 = w.latest_jpeg_since(0.0)
+    assert jpeg1 is not None and ts1 > 0
+
+    # Same frame again: the stream asks "anything newer than ts1?" and the
+    # answer is no. This is what stops duplicates going down the wire.
+    assert w.latest_jpeg_since(ts1) == (None, ts1)
+
+    # The camera produces a new frame; the pipeline still has not run.
+    time.sleep(0.002)
+    w.source.set(grey(160))
+    jpeg2, ts2 = w.latest_jpeg_since(ts1)
+    assert jpeg2 is not None and ts2 > ts1 and jpeg2 != jpeg1
+
+
+def test_boxes_from_the_last_processed_frame_are_drawn_on_the_fresh_one(tmp_path):
+    w = make_worker(tmp_path)
+    w.source.set(grey())
+    plain, _ = w.latest_jpeg_since(0.0)
+
+    # The pipeline finishes a frame with one recognised person in it.
+    w._remember_tracks([resolved_track()])
+
+    # The NEXT camera frame - which the pipeline has not seen - still carries
+    # the box, because a person does not vanish between two frames.
+    time.sleep(0.002)
+    w.source.set(grey())
+    boxed, _ = w.latest_jpeg_since(0.0)
+    assert boxed != plain, "a resolved track should be drawn on the live picture"
+
+
+def test_a_stale_overlay_is_not_drawn(tmp_path):
+    """A stalled pipeline must not leave a box floating over an empty spot."""
+    w = make_worker(tmp_path)
+    w.source.set(grey())
+    plain, _ = w.latest_jpeg_since(0.0)
+
+    w._remember_tracks([resolved_track()])
+    # Pretend the pipeline last ran a while ago.
+    with w._lock:
+        w._overlay_ts = time.time() - 2.0
+
+    time.sleep(0.002)
+    w.source.set(grey())
+    fresh, _ = w.latest_jpeg_since(0.0)
+    assert fresh == plain, "boxes older than a second should not be drawn"
+
+
+def test_only_tracks_matched_in_the_frame_are_drawn(tmp_path):
+    """A track being coasted on misses has no face under it right now."""
+    w = make_worker(tmp_path)
+    missed = resolved_track()
+    missed.misses = 3
+    w._remember_tracks([missed])
+    assert w._overlay == []
+
+    seen = resolved_track()
+    w._remember_tracks([seen])
+    assert len(w._overlay) == 1
+    (box, _color, text) = w._overlay[0]
+    assert box == (30, 30, 90, 100) and text.startswith("Priya (")
+
+
+def test_remembering_tracks_does_not_encode_or_copy(tmp_path):
+    """The whole CPU argument: recording what to draw is a few tuples, and the
+    frame is never touched. A shop PC with no viewer pays nothing."""
+    w = make_worker(tmp_path)
+    tracks = [resolved_track() for _ in range(5)]
+    t0 = time.perf_counter()
+    for _ in range(1000):
+        w._remember_tracks(tracks)
+    per_call_us = (time.perf_counter() - t0) / 1000 * 1e6
+    # A JPEG encode of even a small frame is hundreds of microseconds; this
+    # should be an order of magnitude under that.
+    assert per_call_us < 100, f"remembering tracks took {per_call_us:.0f}us"