Pose: per-person bbox-scaled radius+thickness; green bbox overlay

app.py

@@ -185,18 +185,31 @@ def predict(image: Image.Image, size: str, kpt_thr: float):
     keypoints, scores = _estimate_pose(image_bgr, bboxes, model)
 
     meta = model.pose_metainfo
-    vis_rgb = visualize_keypoints(
-        image=image_rgb,
-        keypoints=keypoints,
-        keypoints_visible=[np.ones(len(s), dtype=bool) for s in scores],
-        keypoint_scores=scores,
-        radius=3,
-        thickness=1,
-        kpt_thr=kpt_thr,
-        skeleton=meta["skeleton_links"],
-        kpt_color=meta["keypoint_colors"],
-        link_color=meta["skeleton_link_colors"],
-    )
+    vis_rgb = image_rgb.copy()
+    for bbox, kpts, scr in zip(bboxes, keypoints, scores):
+        x1, y1, x2, y2 = map(int, bbox[:4])
+        # Scale render sizes with bbox area (per-person), matching v1's recipe.
+        bbox_diag = float(np.sqrt(max(1, x2 - x1) * max(1, y2 - y1)))
+        radius = max(1, int(bbox_diag * 0.006))
+        thickness = max(1, int(bbox_diag * 0.006))
+        bbox_thickness = max(1, thickness // 4)
+
+        # Green bbox first (so skeleton draws on top).
+        cv2.rectangle(vis_rgb, (x1, y1), (x2, y2), (0, 255, 0), bbox_thickness)
+
+        # Skeleton + keypoints for this person only.
+        vis_rgb = visualize_keypoints(
+            image=vis_rgb,
+            keypoints=[kpts],
+            keypoints_visible=[np.ones(len(scr), dtype=bool)],
+            keypoint_scores=[scr],
+            radius=radius,
+            thickness=thickness,
+            kpt_thr=kpt_thr,
+            skeleton=meta["skeleton_links"],
+            kpt_color=meta["keypoint_colors"],
+            link_color=meta["skeleton_link_colors"],
+        )
     vis_pil = Image.fromarray(vis_rgb)
 
     instances = [