scorevision: push artifact

README.md

@@ -1,13 +1,14 @@
 ---
 tags:
 - element_type:detect
-- model:yolov11-small
-- object:vehicle
+- model:yolov11-nano
+- object:person
 manako:
   description: >
-    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
+    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
   prompt_hints: null
   input_payload:
   - name: frame
@@ -16,32 +17,26 @@ manako:
   output_payload:
   - name: detections
     type: detections
-    description: Bounding boxes for detected vehicles
-  evaluation_score: 0.7701
+    description: Bounding boxes for detected persons
+  evaluation_score: 0.5563
   last_benchmark:
-    type: visdrone_val
-    ran_at: 2026-03-25T17:34:00+00:00
+    type: coco_val2017
+    ran_at: '2026-03-25T02:58:57+00:00'
     result_path: null
 ---
 
-# Detect-Vehicle — SN44
+# Detect-Person — SN44
 
-YOLO11s fine-tuned for vehicle detection (car, bus, truck, motorcycle).
+YOLOv11-nano fine-tuned for ground-level CCTV person detection.
 
 | Metric | Value |
 |--------|-------|
-| 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% |
+| 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 |
 
-## Classes
-
-| Output ID | Class |
-|-----------|-------|
-| 0 | car |
-| 1 | bus |
-| 2 | truck |
-| 3 | motorcycle |
+**Training data**: CrowdHuman (15k) + BDD100K (3.2k pedestrians)
+**Validation**: COCO val2017 persons (2,693 images)