scorevision: push artifact

miner.py

@@ -1,5 +1,5 @@
 """
-Detect-Person miner for ScoreVision.
+Detect-vehicle miner for ScoreVision.
 
 Loaded by the TurboVision chute_template from the root of the HF repo.
 Thresholds (imgsz, conf, iou, max_det) are overridable via SN44_* env vars
@@ -8,6 +8,10 @@ so operators can hot-patch without redeploying.
 Contract expected by the chute template:
   * class `Miner(path_hf_repo: Path)`
   * method `predict_batch(batch_images, offset, n_keypoints) -> list[TVFrameResult]`
+
+Vehicle classes filtered from COCO 80-class: car(2), motorcycle(3), bus(5), truck(7).
+cls_id is REMAPPED to 0 in the output because the ScoreVision validator
+treats class-0 as the single target class per element.
 """
 
 from __future__ import annotations
@@ -136,10 +140,18 @@ def _clip_boxes(boxes: np.ndarray, image_size: tuple[int, int]) -> np.ndarray:
 
 
 class Miner:
-    """Detect-Person miner: ONNX Runtime + raw YOLO decode + numpy NMS."""
+    """Detect-vehicle miner: ONNX Runtime + raw YOLO decode + numpy NMS.
+
+    Same yolo11s_fp16 backbone as the Detect-Person miner; inference path is
+    unchanged. The ONLY difference is class filtering: we keep COCO classes
+    {car=2, motorcycle=3, bus=5, truck=7} and merge them into a single
+    emitted class_id=0 per the validator's single-target convention.
+    """
 
     def __init__(self, path_hf_repo: Path) -> None:
-        self.class_names = ["person"]
+        self.class_names = ["vehicle"]
+        # COCO class indices we keep:
+        self.vehicle_coco_ids = (2, 3, 5, 7)  # car, motorcycle, bus, truck
 
         weights_name = os.environ.get("SN44_ONNX_WEIGHTS", _DEFAULT_WEIGHTS)
         weights_path = path_hf_repo / weights_name
@@ -199,21 +211,21 @@ class Miner:
         self.max_aspect_ratio = 8.0
         self.max_box_area_ratio = 0.9
 
-        self.person_cls_idx = 0
-
         print(
-            "Miner ready: "
+            "Vehicle Miner ready: "
             f"imgsz={self.input_height}x{self.input_width}, "
             f"conf={self.conf_thres:.3f}, iou={self.iou_thres:.3f}, "
-            f"max_det={self.max_det}, providers={self.session.get_providers()}"
+            f"max_det={self.max_det}, providers={self.session.get_providers()}, "
+            f"coco_ids={self.vehicle_coco_ids}"
         )
 
     def __repr__(self) -> str:
         return (
-            "DetectPersonMiner("
+            "DetectVehicleMiner("
             f"providers={self.session.get_providers()}, "
             f"imgsz={self.input_height}x{self.input_width}, "
-            f"conf={self.conf_thres}, iou={self.iou_thres})"
+            f"conf={self.conf_thres}, iou={self.iou_thres}, "
+            f"coco_ids={self.vehicle_coco_ids})"
         )
 
     def _preprocess(
@@ -294,13 +306,16 @@ class Miner:
         boxes_xywh = preds[:, :4].astype(np.float32)
         class_scores = preds[:, 4:].astype(np.float32)
 
-        person_scores = class_scores[:, self.person_cls_idx]
-        mask = person_scores >= self.conf_thres
+        # For each detection keep MAX confidence across the vehicle classes.
+        vehicle_idx = np.array(self.vehicle_coco_ids, dtype=np.intp)
+        vehicle_scores_all = class_scores[:, vehicle_idx]
+        vehicle_scores = vehicle_scores_all.max(axis=1)
+        mask = vehicle_scores >= self.conf_thres
         if not np.any(mask):
             return []
 
         boxes_xywh = boxes_xywh[mask]
-        scores = person_scores[mask]
+        scores = vehicle_scores[mask]
 
         boxes = _xywh_to_xyxy(boxes_xywh)