Upload 3 files (#2)

- Upload 3 files (065f8d616842d63259e36a9a04ea18e87a49791f)

README.md

@@ -109,6 +109,22 @@ Every model DetectionBench has trained and evaluated on Global Wheat Head Datase
 | wheat_head       | 67.1 | 27.11 |
 ---
 
+## Per-Country Performance
+
+Domain shift can matter more than the aggregate score above for field deployment, so this evaluates the same test split broken down by the contributing country/institution, using per-image domain metadata (`domain_metadata.json` in this repository) compiled by this project for this stratified evaluation -- not a file shipped with the original GWHD release. Each row below is computed by re-running this exact model's evaluation restricted to that country's images only -- the same mAP definition as the aggregate number above (Ultralytics' `model.val()` for YOLO, Supervision's `MeanAveragePrecision` for RF-DETR), just on a filtered subset, not a separate metric implementation. One test image with no resolvable country in the source metadata (a documented upstream duplicate-filename quirk) is excluded from every row below.
+
+| Country                    | mAP@50          | mAP@50-95         | Test Images        |
+| --------------------------- | --------------- | ----------------- | ------------------ |
+| Australia       | 57.87 | 22.49 | 281 |
+| China       | 88.71 | 43.81 | 200 |
+| Japan       | 71.1 | 38.82 | 60 |
+| Mexico       | 68.54 | 27.99 | 205 |
+| Sudan       | 74.03 | 34.04 | 30 |
+| US       | 64.97 | 23.76 | 605 |
+See `country_breakdown.json` (results) and `domain_metadata.json` (the country/growth-stage mapping used to compute them) in this repository for the raw data behind this table.
+
+---
+
 ## Evaluation Visualizations
 
 This model was evaluated with [Supervision](https://github.com/roboflow/supervision)'s detection metrics, which report mAP/Precision/Recall directly but don't produce PR-curve, F1-curve, or confusion-matrix plot images the way Ultralytics' validator does. See the Performance table above for Precision/Recall/F1 and the per-class table above for the full per-class mAP breakdown.
@@ -178,6 +194,8 @@ detections = model.predict("image.jpg", threshold=0.25)
 checkpoint_best_total.pth
 metrics.csv
 config.json
+country_breakdown.json
+domain_metadata.json
 gwhd_rfdetr-medium_showcase.jpg
 README.md
 ```

country_breakdown.json

@@ -0,0 +1,32 @@
+{
+  "Australia": {
+    "mAP50": 0.5787455439567566,
+    "mAP50_95": 0.22489845752716064,
+    "num_images": 281
+  },
+  "China": {
+    "mAP50": 0.8871119022369385,
+    "mAP50_95": 0.4380679726600647,
+    "num_images": 200
+  },
+  "Japan": {
+    "mAP50": 0.7110053896903992,
+    "mAP50_95": 0.38823771476745605,
+    "num_images": 60
+  },
+  "Mexico": {
+    "mAP50": 0.6854299306869507,
+    "mAP50_95": 0.27985256910324097,
+    "num_images": 205
+  },
+  "Sudan": {
+    "mAP50": 0.7403203248977661,
+    "mAP50_95": 0.34035342931747437,
+    "num_images": 30
+  },
+  "US": {
+    "mAP50": 0.6496736407279968,
+    "mAP50_95": 0.23756475746631622,
+    "num_images": 605
+  }
+}