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).
-Single-pass + NMS. Stretch resize preprocessing.
-Single class: person (cls_id=0).
+Score Vision SN44 — VehicleDetect miner v5 (2026-03-26).
+TTA (3-pass) + inline WBF. Per-class 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
@@ -13,6 +18,94 @@ import onnxruntime as ort
 from numpy import ndarray
 from pydantic import BaseModel
 
+MODEL_TO_OUT: dict[int, int] = {0: 1, 1: 0, 2: 2, 3: 3}
+OUT_NAMES = ["bus", "car", "truck", "motorcycle"]
+NUM_CLASSES = 4
+
+IMG_SIZE = 1280
+CONF_THRESH = 0.55
+IOU_THRESH = 0.45
+WBF_IOU_THR = 0.55
+WBF_SKIP_THR = 0.0001
+TTA_SCALE = 1.2
+
+
+def _wbf(boxes_list: list[np.ndarray], scores_list: list[np.ndarray],
+         labels_list: list[np.ndarray], iou_thr: float = 0.55,
+         skip_box_thr: float = 0.0001) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """Weighted Boxes Fusion (inline, no external dep). Boxes in [0,1] normalized coords."""
+    if not boxes_list:
+        return np.empty((0, 4)), np.empty(0), np.empty(0)
+
+    # Collect all boxes with model index
+    all_boxes, all_scores, all_labels = [], [], []
+    for model_idx, (bx, sc, lb) in enumerate(zip(boxes_list, scores_list, labels_list)):
+        for i in range(len(bx)):
+            if sc[i] < skip_box_thr:
+                continue
+            all_boxes.append(bx[i])
+            all_scores.append(sc[i])
+            all_labels.append(int(lb[i]))
+
+    if not all_boxes:
+        return np.empty((0, 4)), np.empty(0), np.empty(0)
+
+    all_boxes = np.array(all_boxes)
+    all_scores = np.array(all_scores)
+    all_labels = np.array(all_labels, dtype=int)
+
+    n_models = len(boxes_list)
+    fused_boxes, fused_scores, fused_labels = [], [], []
+
+    for cls in np.unique(all_labels):
+        cls_mask = all_labels == cls
+        cls_boxes = all_boxes[cls_mask]
+        cls_scores = all_scores[cls_mask]
+
+        order = cls_scores.argsort()[::-1]
+        cls_boxes = cls_boxes[order]
+        cls_scores = cls_scores[order]
+
+        clusters: list[list[int]] = []
+        cluster_boxes: list[np.ndarray] = []
+
+        for i in range(len(cls_boxes)):
+            matched = -1
+            best_iou = iou_thr
+            for c_idx, c_box in enumerate(cluster_boxes):
+                xx1 = max(cls_boxes[i, 0], c_box[0])
+                yy1 = max(cls_boxes[i, 1], c_box[1])
+                xx2 = min(cls_boxes[i, 2], c_box[2])
+                yy2 = min(cls_boxes[i, 3], c_box[3])
+                inter = max(0, xx2 - xx1) * max(0, yy2 - yy1)
+                a1 = (cls_boxes[i, 2] - cls_boxes[i, 0]) * (cls_boxes[i, 3] - cls_boxes[i, 1])
+                a2 = (c_box[2] - c_box[0]) * (c_box[3] - c_box[1])
+                iou = inter / (a1 + a2 - inter + 1e-9)
+                if iou > best_iou:
+                    best_iou = iou
+                    matched = c_idx
+            if matched >= 0:
+                clusters[matched].append(i)
+                # Update cluster box as weighted average
+                idxs = clusters[matched]
+                weights = cls_scores[idxs]
+                w_sum = weights.sum()
+                cluster_boxes[matched] = (cls_boxes[idxs] * weights[:, None]).sum(0) / w_sum
+            else:
+                clusters.append([i])
+                cluster_boxes.append(cls_boxes[i].copy())
+
+        for c_idx, idxs in enumerate(clusters):
+            weights = cls_scores[idxs]
+            score = weights.sum() / n_models
+            fused_boxes.append(cluster_boxes[c_idx])
+            fused_scores.append(score)
+            fused_labels.append(cls)
+
+    if not fused_boxes:
+        return np.empty((0, 4)), np.empty(0), np.empty(0)
+    return np.array(fused_boxes), np.array(fused_scores), np.array(fused_labels)
+
 
 class BoundingBox(BaseModel):
     x1: int
@@ -32,115 +125,130 @@ 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 v3 NMS 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]:
-        """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)
+        return f"VehicleDetect Miner v5 TTA+WBF session={type(self.session).__name__}"
 
-        if pred.shape[1] < 5:
-            return np.empty((0, 4)), np.empty(0), np.empty(0, dtype=int)
+    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
 
-        boxes = 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)
+    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]:
+        pred = raw[0]
+        if pred.shape[0] < pred.shape[1]:
+            pred = pred.T
+        bboxes_cx = pred[:, :4]
+        cls_scores = pred[:, 4:]
         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]
-
-        if boxes.shape[0] == 0:
+        mask = confs >= self.conf_threshold
+        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)
-
-        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)
-
-        return out_boxes, confs, cls_ids
+        bboxes_cx, confs, cls_ids = bboxes_cx[mask], confs[mask], cls_ids[mask]
+        cx, cy, bw, bh = bboxes_cx[:, 0], bboxes_cx[:, 1], bboxes_cx[:, 2], bboxes_cx[:, 3]
+        x1 = np.clip((cx - bw / 2 - pad_l) / ratio, 0, orig_w)
+        y1 = np.clip((cy - bh / 2 - pad_t) / ratio, 0, orig_h)
+        x2 = np.clip((cx + bw / 2 - pad_l) / ratio, 0, orig_w)
+        y2 = np.clip((cy + bh / 2 - pad_t) / ratio, 0, orig_h)
+        return np.stack([x1, y1, x2, y2], axis=1), confs, cls_ids
 
     def _run_single_pass(self, image_bgr: ndarray) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
         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)
-
-    def _nms_dets(self, boxes: np.ndarray, confs: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
-        """Simple NMS on boxes+confs."""
-        if len(boxes) == 0:
-            return boxes, confs
-        x1, y1, x2, y2 = boxes[:, 0], boxes[:, 1], boxes[:, 2], boxes[:, 3]
-        areas = (x2 - x1) * (y2 - y1)
-        order = confs.argsort()[::-1]
-        keep = []
-        while len(order):
-            i = order[0]
-            keep.append(int(i))
-            xx1 = np.maximum(x1[i], x1[order[1:]])
-            yy1 = np.maximum(y1[i], y1[order[1:]])
-            xx2 = np.minimum(x2[i], x2[order[1:]])
-            yy2 = np.minimum(y2[i], y2[order[1:]])
-            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 self._decode_raw(raw, ratio, pad_l, pad_t, orig_w, orig_h)
 
     def _infer_single(self, image_bgr: ndarray) -> list[BoundingBox]:
         orig_h, orig_w = image_bgr.shape[:2]
-        boxes, confs, cls_ids = self._run_single_pass(image_bgr)
-        if len(boxes) == 0:
+
+        all_boxes, all_scores, all_labels = [], [], []
+
+        def _collect(boxes, confs, cls_ids):
+            if len(boxes) == 0:
+                return
+            out_cls = np.array([MODEL_TO_OUT[int(c)] for c in cls_ids])
+            norm = boxes.copy()
+            norm[:, [0, 2]] /= orig_w
+            norm[:, [1, 3]] /= orig_h
+            norm = np.clip(norm, 0, 1)
+            all_boxes.append(norm)
+            all_scores.append(confs)
+            all_labels.append(out_cls)
+
+        # Pass 1: original
+        _collect(*self._run_single_pass(image_bgr))
+
+        # Pass 2: horizontal flip
+        flipped = cv2.flip(image_bgr, 1)
+        boxes_f, confs_f, cls_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]
+        _collect(boxes_f, confs_f, cls_f)
+
+        # Pass 3: 1.2x scale center crop
+        sh, sw = int(orig_h * TTA_SCALE), int(orig_w * TTA_SCALE)
+        scaled = cv2.resize(image_bgr, (sw, sh), interpolation=cv2.INTER_LINEAR)
+        yo, xo = (sh - orig_h) // 2, (sw - orig_w) // 2
+        cropped = scaled[yo:yo + orig_h, xo:xo + orig_w]
+        boxes_s, confs_s, cls_s = self._run_single_pass(cropped)
+        if len(boxes_s):
+            boxes_s[:, 0] = (boxes_s[:, 0] + xo) / TTA_SCALE
+            boxes_s[:, 1] = (boxes_s[:, 1] + yo) / TTA_SCALE
+            boxes_s[:, 2] = (boxes_s[:, 2] + xo) / TTA_SCALE
+            boxes_s[:, 3] = (boxes_s[:, 3] + yo) / TTA_SCALE
+            boxes_s = np.clip(boxes_s, 0, [[orig_w, orig_h, orig_w, orig_h]])
+        _collect(boxes_s, confs_s, cls_s)
+
+        if not all_boxes:
             return []
 
-        boxes, confs = self._nms_dets(boxes, confs)
-        # cls_ids got filtered by _run_single_pass but not by _nms_dets, re-derive
-        # Actually _nms_dets only returns boxes/confs, cls_ids are all 0 for person
-        out_boxes: list[BoundingBox] = []
-        for i in range(len(boxes)):
-            out_boxes.append(BoundingBox(
-                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=0,
-                conf=max(0.0, min(1.0, float(confs[i]))),
+        fused_boxes, fused_scores, fused_labels = _wbf(
+            all_boxes, all_scores, all_labels,
+            iou_thr=WBF_IOU_THR, skip_box_thr=WBF_SKIP_THR,
+        )
+        if len(fused_boxes) == 0:
+            return []
+
+        # Denormalize
+        fused_boxes[:, [0, 2]] *= orig_w
+        fused_boxes[:, [1, 3]] *= orig_h
+
+        out: list[BoundingBox] = []
+        for i in range(len(fused_boxes)):
+            b = fused_boxes[i]
+            out.append(BoundingBox(
+                x1=max(0, min(orig_w, math.floor(b[0]))),
+                y1=max(0, min(orig_h, math.floor(b[1]))),
+                x2=max(0, min(orig_w, math.ceil(b[2]))),
+                y2=max(0, min(orig_h, math.ceil(b[3]))),
+                cls_id=int(fused_labels[i]),
+                conf=max(0.0, min(1.0, float(fused_scores[i]))),
             ))
-        return out_boxes
+        return out
 
     def predict_batch(
         self,
@@ -152,11 +260,7 @@ 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