6.4

code/video_clip_cls/scripts/pipeline/segment_processor.py

@@ -41,6 +41,32 @@ from ultralytics import YOLO  # noqa: E402
 from pipeline.hand_roi_merge import HandRoiGrouper, two_largest_hands, union_xyxy  # noqa: E402
 
 # 与 run_haocai_actionformer_consumables_e2e 段内失败 return 文案一致，供 Phase2 重试判断
+
+
+def _detect_hands_on_frame(
+    det: Any,
+    fr: np.ndarray,
+    det_conf: float,
+    imgsz_det: int,
+    predict_kw: dict[str, Any] | None,
+) -> list[list[float]]:
+    try:
+        from hand_detector import detect_hands_xyxy
+
+        return detect_hands_xyxy(
+            det,
+            fr,
+            det_conf=det_conf,
+            imgsz_det=imgsz_det,
+            predict_kw=predict_kw,
+        )
+    except ImportError:
+        pred_kw = dict(predict_kw or {})
+        r0 = det.predict(
+            fr, conf=det_conf, imgsz=imgsz_det, verbose=False, **pred_kw
+        )[0]
+        return collect_hand_boxes(det, r0.boxes) if r0.boxes else []
+
 REASON_NO_VALID_HAOCAI_FRAMES = "（无有效耗材帧：好帧/白名单/耗材置信度未全部满足）"
 # 推流 / TSV 离线（无好坏帧门控）
 REASON_NO_VALID_HAOCAI_FRAMES_STREAM = "（无有效耗材帧：白名单/耗材置信度未满足）"
@@ -312,8 +338,9 @@ def process_segment_multi_hand_tear(
         if frame_stride > 1 and (frames_read_in_segment - 1) % frame_stride != 0:
             return
 
-        r0 = det.predict(fr, conf=det_conf, imgsz=imgsz_det, verbose=False, **fg.predict_kw)[0]
-        hands = collect_hand_boxes(det, r0.boxes) if r0.boxes else []
+        hands = _detect_hands_on_frame(
+            det, fr, det_conf, imgsz_det, fg.predict_kw
+        )
         if len(hands) < 2:
             return
 
@@ -602,8 +629,7 @@ def process_segment_haocai_from_frames(
         if frame_stride > 1 and (frames_in_segment - 1) % frame_stride != 0:
             return
 
-        r0 = det.predict(fr, conf=det_conf, imgsz=imgsz_det, verbose=False, **pred_kw)[0]
-        hands = collect_hand_boxes(det, r0.boxes) if r0.boxes else []
+        hands = _detect_hands_on_frame(det, fr, det_conf, imgsz_det, pred_kw)
         crop = _crop_two_hands_union(fr, hands, pad_ratio)
         if crop is None:
             return
@@ -702,8 +728,7 @@ def process_segment_haocai_from_cap(
             return
 
         img_h, img_w = fr.shape[:2]
-        r0 = det.predict(fr, conf=det_conf, imgsz=imgsz_det, verbose=False, **pred_kw)[0]
-        hands = collect_hand_boxes(det, r0.boxes) if r0.boxes else []
+        hands = _detect_hands_on_frame(det, fr, det_conf, imgsz_det, pred_kw)
         crop = _crop_two_hands_union(fr, hands, pad_ratio)
         if crop is None:
             return