[Refactor] Fix bugs of browsing transformed training images (open-mml…

…ab#1591)

* fix bugs of browsing transformed img

* fix unittest

mmpose/datasets/transforms/formatting.py

@@ -77,7 +77,8 @@ class PackPoseInputs(BaseTransform):
         'bbox_scale': 'bbox_scales',
         'bbox_score': 'bbox_scores',
         'keypoints': 'keypoints',
-        'keypoints_visible': 'keypoints_visible'
+        'keypoints_visible': 'keypoints_visible',
+        'transformed_keypoints': 'transformed_keypoints'
     }
 
     label_mapping_table = {
@@ -94,8 +95,10 @@ class PackPoseInputs(BaseTransform):
     def __init__(self,
                  meta_keys=('id', 'img_id', 'img_path', 'ori_shape',
                             'img_shape', 'input_size', 'flip',
-                            'flip_direction', 'flip_indices')):
+                            'flip_direction', 'flip_indices'),
+                 pack_transformed=False):
         self.meta_keys = meta_keys
+        self.pack_transformed = pack_transformed
 
     def transform(self, results: dict) -> dict:
         """Method to pack the input data.
@@ -122,6 +125,9 @@ def transform(self, results: dict) -> dict:
         for key, packed_key in self.instance_mapping_table.items():
             if key in results:
                 gt_instances.set_field(results[key], packed_key)
+        if not self.pack_transformed:
+            if 'transformed_keypoints' in gt_instances:
+                del gt_instances['transformed_keypoints']
         data_sample.gt_instances = gt_instances
 
         # pack instance labels

mmpose/engine/hooks/visualization_hook.py

@@ -4,6 +4,7 @@
 from typing import Optional, Sequence
 
 import mmcv
+import mmengine
 from mmengine.hooks import Hook
 from mmengine.runner import Runner
 from mmengine.visualization import Visualizer
@@ -131,7 +132,7 @@ def after_test_iter(self, runner: Runner, batch_idx: int,
             mmcv.mkdir_or_exist(self.test_out_dir)
 
         if self.file_client is None:
-            self.file_client = mmcv.FileClient(**self.file_client_args)
+            self.file_client = mmengine.FileClient(**self.file_client_args)
 
         for input_data, output in zip(data_batch, outputs):
             self._test_index += 1

mmpose/structures/__init__.py

@@ -1,5 +1,13 @@
 # Copyright (c) OpenMMLab. All rights reserved.
+from .bbox import (bbox_cs2xywh, bbox_cs2xyxy, bbox_xywh2cs, bbox_xywh2xyxy,
+                   bbox_xyxy2cs, bbox_xyxy2xywh, flip_bbox,
+                   get_udp_warp_matrix, get_warp_matrix)
+from .keypoint import flip_keypoints
 from .multilevel_pixel_data import MultilevelPixelData
 from .pose_data_sample import PoseDataSample
 
-__all__ = ['PoseDataSample', 'MultilevelPixelData']
+__all__ = [
+    'PoseDataSample', 'MultilevelPixelData', 'bbox_cs2xywh', 'bbox_cs2xyxy',
+    'bbox_xywh2cs', 'bbox_xywh2xyxy', 'bbox_xyxy2cs', 'bbox_xyxy2xywh',
+    'flip_bbox', 'get_udp_warp_matrix', 'get_warp_matrix', 'flip_keypoints'
+]

mmpose/visualization/local_visualizer.py

@@ -158,7 +158,8 @@ def _draw_instances_kpts(self,
         img_h, img_w, _ = image.shape
 
         if 'keypoints' in instances:
-            keypoints = instances.keypoints
+            keypoints = instances.get('transformed_keypoints',
+                                      instances.keypoints)
 
             if 'scores' in instances and self.show_keypoint_weight:
                 scores = instances.scores

tests/test_datasets/test_transforms/test_formatting.py

@@ -60,11 +60,19 @@ def setUp(self):
             np.random.randint(0, 100, (1, 17, 2)).astype(np.float32),
             'keypoint_y_labels':
             np.random.randint(0, 100, (1, 17, 2)).astype(np.float32),
+            'transformed_keypoints':
+            np.random.randint(0, 100, (1, 17, 2)).astype(np.float32),
         }
         self.meta_keys = ('img_id', 'img_path', 'ori_shape', 'img_shape',
                           'scale_factor', 'flip', 'flip_direction')
 
     def test_transform(self):
+        transform = PackPoseInputs(
+            meta_keys=self.meta_keys, pack_transformed=True)
+        results = transform(copy.deepcopy(self.results_topdown))
+        self.assertIn('transformed_keypoints',
+                      results['data_sample'].gt_instances)
+
         transform = PackPoseInputs(meta_keys=self.meta_keys)
         results = transform(copy.deepcopy(self.results_topdown))
         self.assertIn('inputs', results)
@@ -78,6 +86,8 @@ def test_transform(self):
         self.assertEqual(len(results['data_sample'].gt_instances), 1)
         self.assertIsInstance(results['data_sample'].gt_fields.heatmaps,
                               torch.Tensor)
+        self.assertNotIn('transformed_keypoints',
+                         results['data_sample'].gt_instances)
 
         # test when results['img'] is sequence of frames
         results = copy.deepcopy(self.results_topdown)

tools/misc/browse_dataset.py

@@ -37,7 +37,7 @@ def parse_args():
         help='the interval of show (s)')
     parser.add_argument(
         '--mode',
-        default='original',
+        default='transformed',
         type=str,
         choices=['original', 'transformed'],
         help='display mode; display original pictures or transformed pictures'
@@ -90,6 +90,10 @@ def main():
     if 'bbox_file' in cfg[f'{args.phase}_dataloader'].dataset:
         cfg[f'{args.phase}_dataloader'].dataset.bbox_file = None
 
+    # pack transformed keypoints for visualization
+    cfg[f'{args.phase}_dataloader'].dataset.pipeline[
+        -1].pack_transformed = True
+
     dataset = build_from_cfg(cfg[f'{args.phase}_dataloader'].dataset, DATASETS)
 
     visualizer = VISUALIZERS.build(cfg.visualizer)