scorevision: push artifact

README.md

@@ -1,14 +1,13 @@
 ---
 tags:
 - element_type:detect
-- model:yolov11-nano
-- object:person
+- model:yolov11-small
+- object:vehicle
 manako:
   description: >
-    YOLOv11-nano fine-tuned for ground-level CCTV person detection on SN44.
-    Trained on CrowdHuman (15k, dense crowds) + BDD100K street pedestrians.
-    Conf threshold raised to 0.35 to minimise false positives.
-  source: meaculpitt/Detect-Person
+    YOLO11s vehicle detector fine-tuned on COCO vehicles + BDD100K + VisDrone.
+    FP16 ONNX, 1280x1280 input. Trained R6: 59,870 images, 50 epochs.
+  source: meaculpitt/Detect-Vehicle
   prompt_hints: null
   input_payload:
   - name: frame
@@ -17,26 +16,32 @@ manako:
   output_payload:
   - name: detections
     type: detections
-    description: Bounding boxes for detected persons
-  evaluation_score: 0.5563
+    description: Bounding boxes for detected vehicles
+  evaluation_score: 0.7701
   last_benchmark:
-    type: coco_val2017
-    ran_at: '2026-03-25T02:58:57+00:00'
+    type: visdrone_val
+    ran_at: 2026-03-25T17:34:00+00:00
     result_path: null
 ---
 
-# Detect-Person — SN44
+# Detect-Vehicle — SN44
 
-YOLOv11-nano fine-tuned for ground-level CCTV person detection.
+YOLO11s fine-tuned for vehicle detection (car, bus, truck, motorcycle).
 
 | Metric | Value |
 |--------|-------|
-| mAP@50 (COCO val2017) | 55.63% |
-| Precision (conf=0.35) | 56.86% |
-| Recall | 50.67% |
-| Baseline to beat | 37.55% |
-| Model size | 5.6 MB |
-| Input size | 1280×1280 |
+| mAP@50 | 77.01% |
+| Model | YOLO11s (FP16 ONNX) |
+| Input size | 1280x1280 |
+| Model size | 19.2 MB |
+| Training data | COCO vehicles + BDD100K + VisDrone (59,870 images) |
+| Baseline to beat | 40.72% |
 
-**Training data**: CrowdHuman (15k) + BDD100K (3.2k pedestrians)
-**Validation**: COCO val2017 persons (2,693 images)
+## Classes
+
+| Output ID | Class |
+|-----------|-------|
+| 0 | car |
+| 1 | bus |
+| 2 | truck |
+| 3 | motorcycle |

class_names.txt

@@ -1 +1,80 @@
 person
+bicycle
+car
+motorcycle
+airplane
+bus
+train
+truck
+boat
+traffic light
+fire hydrant
+stop sign
+parking meter
+bench
+bird
+cat
+dog
+horse
+sheep
+cow
+elephant
+bear
+zebra
+giraffe
+backpack
+umbrella
+handbag
+tie
+suitcase
+frisbee
+skis
+snowboard
+sports ball
+kite
+baseball bat
+baseball glove
+skateboard
+surfboard
+tennis racket
+bottle
+wine glass
+cup
+fork
+knife
+spoon
+bowl
+banana
+apple
+sandwich
+orange
+broccoli
+carrot
+hot dog
+pizza
+donut
+cake
+chair
+couch
+potted plant
+bed
+dining table
+toilet
+tv
+laptop
+mouse
+remote
+keyboard
+cell phone
+microwave
+oven
+toaster
+sink
+refrigerator
+book
+clock
+vase
+scissors
+teddy bear
+hair drier
+toothbrush

miner.py

@@ -1,7 +1,12 @@
 """
-Score Vision SN44 — DetectPerson miner. v2 (2026-03-26).
-TTA (3 augmentations) + WBF fusion. Stretch resize preprocessing.
-Single class: person (cls_id=0).
+Score Vision SN44 — VehicleDetect miner. v3 (2026-03-26).
+Single-pass + NMS. Letterbox preprocessing.
+
+Model: YOLO11s ONNX, 4 classes trained as:
+  0 = car, 1 = bus, 2 = truck, 3 = motorcycle
+
+Official submission order (remapped in MODEL_TO_OUT):
+  0 = bus, 1 = car, 2 = truck, 3 = motorcycle
 """
 
 from pathlib import Path
@@ -12,11 +17,14 @@ import numpy as np
 import onnxruntime as ort
 from numpy import ndarray
 from pydantic import BaseModel
-from ensemble_boxes import weighted_boxes_fusion
+# ── Model class index → submission class index ───────────────────────────────
+MODEL_TO_OUT: dict[int, int] = {0: 1, 1: 0, 2: 2, 3: 3}
+OUT_TO_MODEL: dict[int, int] = {v: k for k, v in MODEL_TO_OUT.items()}
+OUT_NAMES = ["bus", "car", "truck", "motorcycle"]
 
-WBF_IOU_THR = 0.55
-WBF_SKIP_THR = 0.0001
-TTA_SCALE = 1.2
+IMG_SIZE = 1280
+CONF_THRESH = 0.55
+IOU_THRESH = 0.45
 
 
 class BoundingBox(BaseModel):
@@ -37,79 +45,83 @@ class TVFrameResult(BaseModel):
 class Miner:
     def __init__(self, path_hf_repo: Path) -> None:
         self.path_hf_repo = path_hf_repo
-        self.class_names = ['person']
         self.session = ort.InferenceSession(
             str(path_hf_repo / "weights.onnx"),
             providers=["CUDAExecutionProvider", "CPUExecutionProvider"],
         )
         self.input_name = self.session.get_inputs()[0].name
-        input_shape = self.session.get_inputs()[0].shape
-        self.input_h = int(input_shape[2])
-        self.input_w = int(input_shape[3])
-        self.conf_threshold = 0.50
-        self.iou_threshold = 0.45
+        self.conf_threshold = CONF_THRESH
+        self.iou_threshold = IOU_THRESH
 
     def __repr__(self) -> str:
-        return f"DetectPerson Miner TTA+WBF session={type(self.session).__name__}"
-
-    def _preprocess(self, image_bgr: ndarray) -> tuple[np.ndarray, tuple[int, int]]:
-        h, w = image_bgr.shape[:2]
-        rgb = cv2.cvtColor(image_bgr, cv2.COLOR_BGR2RGB)
-        resized = cv2.resize(rgb, (self.input_w, self.input_h))
-        x = resized.astype(np.float32) / 255.0
-        x = np.transpose(x, (2, 0, 1))[None, ...]
-        return x, (h, w)
-
-    def _decode_raw(self, raw: np.ndarray, orig_h: int, orig_w: int
-                    ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        return f"VehicleDetect Miner v4 NMS session={type(self.session).__name__}"
+
+    def _letterbox(self, img: ndarray) -> tuple[np.ndarray, float, int, int]:
+        h, w = img.shape[:2]
+        r = min(IMG_SIZE / h, IMG_SIZE / w)
+        new_w, new_h = int(round(w * r)), int(round(h * r))
+        img_r = cv2.resize(img, (new_w, new_h), interpolation=cv2.INTER_LINEAR)
+        dw, dh = IMG_SIZE - new_w, IMG_SIZE - new_h
+        pad_l, pad_t = dw // 2, dh // 2
+        img_p = cv2.copyMakeBorder(
+            img_r, pad_t, dh - pad_t, pad_l, dw - pad_l,
+            cv2.BORDER_CONSTANT, value=(114, 114, 114),
+        )
+        return img_p, r, pad_l, pad_t
+
+    def _preprocess(self, image_bgr: ndarray) -> tuple[np.ndarray, float, int, int]:
+        img_p, ratio, pad_l, pad_t = self._letterbox(image_bgr)
+        img_rgb = cv2.cvtColor(img_p, cv2.COLOR_BGR2RGB)
+        inp = img_rgb.astype(np.float32) / 255.0
+        inp = np.ascontiguousarray(inp.transpose(2, 0, 1)[np.newaxis])
+        return inp, ratio, pad_l, pad_t
+
+    def _decode_raw(self, raw: np.ndarray, ratio: float, pad_l: int, pad_t: int,
+                    orig_w: int, orig_h: int) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
         """Decode ONNX output to (boxes_xyxy, confs, cls_ids) in original image coords."""
         pred = raw[0]
-        if pred.ndim != 2:
-            return np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int)
         if pred.shape[0] < pred.shape[1]:
-            pred = pred.transpose(1, 0)
-
-        if pred.shape[1] < 5:
-            return np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int)
+            pred = pred.T
 
-        boxes = pred[:, :4]
+        bboxes_cx = pred[:, :4]
         cls_scores = pred[:, 4:]
-        if cls_scores.shape[1] == 0:
-            return np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int)
 
         cls_ids = np.argmax(cls_scores, axis=1)
         confs = np.max(cls_scores, axis=1)
-        keep = confs >= self.conf_threshold
-
-        boxes = boxes[keep]
-        confs = confs[keep]
-        cls_ids = cls_ids[keep]
+        mask = confs >= self.conf_threshold
 
-        if boxes.shape[0] == 0:
+        if not mask.any():
             return np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int)
 
-        sx = orig_w / float(self.input_w)
-        sy = orig_h / float(self.input_h)
+        bboxes_cx = bboxes_cx[mask]
+        confs = confs[mask]
+        cls_ids = cls_ids[mask]
 
-        cx, cy, bw, bh = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
-        x1 = np.clip((cx - bw / 2) * sx, 0, orig_w)
-        y1 = np.clip((cy - bh / 2) * sy, 0, orig_h)
-        x2 = np.clip((cx + bw / 2) * sx, 0, orig_w)
-        y2 = np.clip((cy + bh / 2) * sy, 0, orig_h)
-        out_boxes = np.stack([x1, y1, x2, y2], axis=1)
+        cx, cy, bw, bh = bboxes_cx[:, 0], bboxes_cx[:, 1], bboxes_cx[:, 2], bboxes_cx[:, 3]
+        lx1 = cx - bw / 2
+        ly1 = cy - bh / 2
+        lx2 = cx + bw / 2
+        ly2 = cy + bh / 2
 
-        return out_boxes, confs, cls_ids
+        x1 = np.clip((lx1 - pad_l) / ratio, 0, orig_w)
+        y1 = np.clip((ly1 - pad_t) / ratio, 0, orig_h)
+        x2 = np.clip((lx2 - pad_l) / ratio, 0, orig_w)
+        y2 = np.clip((ly2 - pad_t) / ratio, 0, orig_h)
+        boxes = np.stack([x1, y1, x2, y2], axis=1)
+
+        return boxes, confs, cls_ids
 
     def _run_single_pass(self, image_bgr: ndarray) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """Run one inference pass, return (boxes_xyxy, confs, cls_ids) in original coords."""
         orig_h, orig_w = image_bgr.shape[:2]
-        inp, _ = self._preprocess(image_bgr)
+        inp, ratio, pad_l, pad_t = self._preprocess(image_bgr)
         raw = self.session.run(None, {self.input_name: inp})[0]
-        return self._decode_raw(raw, orig_h, orig_w)
+        return self._decode_raw(raw, ratio, pad_l, pad_t, orig_w, orig_h)
 
-    def _nms_dets(self, boxes: np.ndarray, confs: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
-        """Simple NMS on boxes+confs."""
+    def _nms(self, boxes: np.ndarray, confs: np.ndarray, cls_ids: np.ndarray
+             ) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
         if len(boxes) == 0:
-            return boxes, confs
+            return boxes, confs, cls_ids
         x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
         areas = (x2 - x1) * (y2 - y1)
         order = confs.argsort()[::-1]
@@ -124,86 +136,25 @@ class Miner:
             inter = np.maximum(0, xx2 - xx1) * np.maximum(0, yy2 - yy1)
             iou = inter / (areas[i] + areas[order[1:]] - inter + 1e-7)
             order = order[1:][iou <= self.iou_threshold]
-        return boxes[keep], confs[keep]
+        return boxes[keep], confs[keep], cls_ids[keep]
 
     def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:
         orig_h, orig_w = image_bgr.shape[:2]
-
-        all_boxes_list = []
-        all_scores_list = []
-        all_labels_list = []
-
-        # ── TTA pass 1: original ──
         boxes, confs, cls_ids = self._run_single_pass(image_bgr)
-        if len(boxes):
-            norm_boxes = boxes.copy()
-            norm_boxes[:, [0, 2]] /= orig_w
-            norm_boxes[:, [1, 3]] /= orig_h
-            norm_boxes = np.clip(norm_boxes, 0, 1)
-            all_boxes_list.append(norm_boxes)
-            all_scores_list.append(confs)
-            all_labels_list.append(cls_ids.astype(float))
-
-        # ── TTA pass 2: horizontal flip ──
-        flipped = cv2.flip(image_bgr, 1)
-        boxes_f, confs_f, cls_ids_f = self._run_single_pass(flipped)
-        if len(boxes_f):
-            boxes_f[:, 0], boxes_f[:, 2] = orig_w - boxes_f[:, 2], orig_w - boxes_f[:, 0]
-            norm_boxes_f = boxes_f.copy()
-            norm_boxes_f[:, [0, 2]] /= orig_w
-            norm_boxes_f[:, [1, 3]] /= orig_h
-            norm_boxes_f = np.clip(norm_boxes_f, 0, 1)
-            all_boxes_list.append(norm_boxes_f)
-            all_scores_list.append(confs_f)
-            all_labels_list.append(cls_ids_f.astype(float))
-
-        # ── TTA pass 3: scale 1.2x center crop ──
-        scaled_h, scaled_w = int(orig_h * TTA_SCALE), int(orig_w * TTA_SCALE)
-        scaled = cv2.resize(image_bgr, (scaled_w, scaled_h), interpolation=cv2.INTER_LINEAR)
-        y_off = (scaled_h - orig_h) // 2
-        x_off = (scaled_w - orig_w) // 2
-        cropped = scaled[y_off:y_off + orig_h, x_off:x_off + orig_w]
-        boxes_s, confs_s, cls_ids_s = self._run_single_pass(cropped)
-        if len(boxes_s):
-            boxes_s[:, 0] = (boxes_s[:, 0] + x_off) / TTA_SCALE
-            boxes_s[:, 1] = (boxes_s[:, 1] + y_off) / TTA_SCALE
-            boxes_s[:, 2] = (boxes_s[:, 2] + x_off) / TTA_SCALE
-            boxes_s[:, 3] = (boxes_s[:, 3] + y_off) / TTA_SCALE
-            boxes_s = np.clip(boxes_s, 0, [[orig_w, orig_h, orig_w, orig_h]])
-            norm_boxes_s = boxes_s.copy()
-            norm_boxes_s[:, [0, 2]] /= orig_w
-            norm_boxes_s[:, [1, 3]] /= orig_h
-            norm_boxes_s = np.clip(norm_boxes_s, 0, 1)
-            all_boxes_list.append(norm_boxes_s)
-            all_scores_list.append(confs_s)
-            all_labels_list.append(cls_ids_s.astype(float))
-
-        if not all_boxes_list:
-            return []
-
-        # ── WBF fusion ──
-        fused_boxes, fused_scores, fused_labels = weighted_boxes_fusion(
-            all_boxes_list, all_scores_list, all_labels_list,
-            iou_thr=WBF_IOU_THR, skip_box_thr=WBF_SKIP_THR,
-        )
-
-        if len(fused_boxes) == 0:
+        if len(boxes) == 0:
             return []
 
-        # Denormalize
-        fused_boxes[:, [0, 2]] *= orig_w
-        fused_boxes[:, [1, 3]] *= orig_h
+        boxes, confs, cls_ids = self._nms(boxes, confs, cls_ids)
 
         out_boxes: list[BoundingBox] = []
-        for i in range(len(fused_boxes)):
-            box = fused_boxes[i]
+        for i in range(len(boxes)):
             out_boxes.append(BoundingBox(
-                x1=max(0, min(orig_w, math.floor(box[0]))),
-                y1=max(0, min(orig_h, math.floor(box[1]))),
-                x2=max(0, min(orig_w, math.ceil(box[2]))),
-                y2=max(0, min(orig_h, math.ceil(box[3]))),
-                cls_id=int(fused_labels[i]),
-                conf=max(0.0, min(1.0, float(fused_scores[i]))),
+                x1=max(0, min(orig_w, math.floor(boxes[i, 0]))),
+                y1=max(0, min(orig_h, math.floor(boxes[i, 1]))),
+                x2=max(0, min(orig_w, math.ceil(boxes[i, 2]))),
+                y2=max(0, min(orig_h, math.ceil(boxes[i, 3]))),
+                cls_id=MODEL_TO_OUT[int(cls_ids[i])],
+                conf=max(0.0, min(1.0, float(confs[i]))),
             ))
         return out_boxes
 
@@ -217,11 +168,9 @@ class Miner:
         for idx, image in enumerate(batch_images):
             boxes = self._infer_single(image)
             keypoints = [(0, 0) for _ in range(max(0, int(n_keypoints)))]
-            results.append(
-                TVFrameResult(
-                    frame_id=offset + idx,
-                    boxes=boxes,
-                    keypoints=keypoints,
-                )
-            )
+            results.append(TVFrameResult(
+                frame_id=offset + idx,
+                boxes=boxes,
+                keypoints=keypoints,
+            ))
         return results

model_type.json

@@ -1 +1 @@
-{"task_type": "object-detection", "model_type": "yolov11-nano", "deploy": "2026-03-26T07:46Z"}
+{"task_type": "object-detection", "model_type": "yolov11-small", "deploy": "2026-03-26T07:43Z"}

weights.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f32ed65b9024a69693f675d494c7fc813a964766c54b241464a463377342da60
-size 5607862
+oid sha256:e3916408ec21f8c94358c18914f922814770b78557e52fe17ff7a9ee74339a5a
+size 19272252