Update app.py

app.py

@@ -5,6 +5,10 @@
 # - (NEW) Per-employee live timers on video overlays: Working and Idle counters (mm:ss)
 # - (NEW) Employees view shows the same live Working/Idle seconds
 # - (Patch I–K) Real-time streaming pace: higher default FPS, up to 60 FPS slider, process every frame when possible, time-synced display (no fixed sleeps).
+# - (Patch M) Real-time catch-up reader (clip plays in wall time; drop backlog if behind)
+# - (Patch O) Downscale only for inference (rescale boxes for draw)
+# - (Patch P) Run InsightFace every ~5 processed frames (reuse embeddings in between)
+# - (Patch Q) Update progress bar every ~10 frames
 
 import os, json, math, tempfile, time
 from datetime import datetime, date
@@ -683,9 +687,20 @@ with tab_live:
         last_display_t = time.perf_counter()
         target_dt = 1.0 / float(src_fps if src_fps and src_fps > 0 else 25.0)
 
+        # Patch M: wall-clock anchor (real-time catch-up)
+        wall_t0 = time.perf_counter()
+
         while cap.isOpened() and frame_no < max_frames:
+            # ---- Patch M: real-time catch-up reader ----
+            elapsed = time.perf_counter() - wall_t0
+            target_idx = int(elapsed * (src_fps if src_fps and src_fps > 0 else 25.0))
+            cur_idx = int(cap.get(cv2.CAP_PROP_POS_FRAMES))
+            if target_idx > cur_idx + 1:
+                cap.set(cv2.CAP_PROP_POS_FRAMES, target_idx)
+
             ret, frame = cap.read()
-            if not ret: break
+            if not ret:
+                break
             pos = int(cap.get(cv2.CAP_PROP_POS_FRAMES))
             if pos % step != 0:
                 frame_no += 1
@@ -696,6 +711,21 @@ with tab_live:
                         last_display_t = time.perf_counter()
                 continue
 
+            # ---- Patch O: downscale only for inference; rescale boxes for draw
+            H0, W0 = frame.shape[:2]
+            infer_scale = 1.0
+            if W0 > 1280:
+                infer_scale = 1280.0 / W0
+                infer_frame = cv2.resize(frame, (int(W0*infer_scale), int(H0*infer_scale)), interpolation=cv2.INTER_AREA)
+            else:
+                infer_frame = frame
+
+            def _inv(b):
+                if infer_scale == 1.0: return b
+                x1,y1,x2,y2 = b
+                inv = 1.0 / infer_scale
+                return (int(x1*inv), int(y1*inv), int(x2*inv), int(y2*inv))
+
             do_process = True
             gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
             if prev_gray is not None and skip_near_identical:
@@ -712,15 +742,27 @@ with tab_live:
                 frame_no += 1
                 continue
 
-            dets = run_yolo_on_frame(model, frame, conf_thres)
-            persons = [d for d in dets if d.cls == "person"]
-            phones  = [d for d in dets if d.cls in ("cell phone","mobile phone","phone")]
+            # YOLO on infer_frame, rescale boxes back
+            dets = run_yolo_on_frame(model, infer_frame, conf_thres)
+            persons = [DetBox(d.cls, d.conf, _inv(d.box)) for d in dets if d.cls == "person"]
+            phones  = [DetBox(d.cls, d.conf, _inv(d.box)) for d in dets if d.cls in ("cell phone","mobile phone","phone")]
 
             vis = blur_faces_if_needed(frame, face_cascade, st.session_state.privacy_blur)
-            rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
 
-            face_pairs = get_face_pairs_insight(rgb)
-            face_boxes_only = [b for (b,_) in face_pairs]
+            # ---- Patch P: run InsightFace every ~5 processed frames; reuse otherwise
+            if 'last_face_pairs' not in st.session_state: st.session_state.last_face_pairs = []
+            if 'last_face_update_pos' not in st.session_state: st.session_state.last_face_update_pos = -999
+            if (pos - st.session_state.last_face_update_pos) >= 5:
+                rgb_infer = cv2.cvtColor(infer_frame, cv2.COLOR_BGR2RGB)
+                face_pairs = get_face_pairs_insight(rgb_infer)
+                st.session_state.last_face_pairs = face_pairs
+                st.session_state.last_face_update_pos = pos
+            else:
+                face_pairs = st.session_state.last_face_pairs
+            face_boxes_only = [_inv(b) for (b, _) in face_pairs]
+
+            # (legacy var kept; harmless)
+            rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
 
             for idx, p in enumerate(persons):
                 # motion / base activity
@@ -732,7 +774,7 @@ with tab_live:
                 # identify person
                 name = None
                 for (fbox, emb) in face_pairs:
-                    fx1, fy1, fx2, fy2 = fbox
+                    fx1, fy1, fx2, fy2 = _inv(fbox)
                     cx = (fx1+fx2)/2; cy = (fy1+fy2)/2
                     if (p.box[0] <= cx <= p.box[2]) and (p.box[1] <= cy <= p.box[3]):
                         name = recognize_name_from_embedding(emb, st.session_state.face_db)
@@ -856,7 +898,9 @@ with tab_live:
                     last_display_t = time.perf_counter()
 
             frame_no += 1
-            prog.progress(min(1.0, frame_no/max_frames))
+            # Patch Q: throttle progress updates
+            if frame_no % 10 == 0:
+                prog.progress(min(1.0, frame_no/max_frames))
 
         cap.release(); os.remove(tmp_path); prog.progress(1.0)
         st.success(f"Tile {tile_index+1}: processed {frame_no} frames")