[Fix] Refine h36m data loading and add head_size to PackPoseInputs (#2735)

Author: Tau-J

mmpose/datasets/datasets/base/base_mocap_dataset.py

@@ -361,9 +361,10 @@ def _load_annotations(self) -> Tuple[List[dict], List[dict]]:
 
             instance_list.append(instance_info)
 
-        for idx, imgname in enumerate(img_names):
-            img_info = self.get_img_info(idx, imgname)
-            image_list.append(img_info)
+        if self.data_mode == 'bottomup':
+            for idx, imgname in enumerate(img_names):
+                img_info = self.get_img_info(idx, imgname)
+                image_list.append(img_info)
 
         return instance_list, image_list
 

mmpose/datasets/transforms/formatting.py

@@ -114,6 +114,9 @@ class PackPoseInputs(BaseTransform):
         # using gt_instances.bbox_scales. To unsure correspondence with
         # previous version, this key is preserved here.
         bbox_scale='bbox_scales',
+        # `head_size` is used for computing MpiiPCKAccuracy metric,
+        # namely, PCKh
+        head_size='head_size',
     )
 
     # items in `field_mapping_table` will be packed into