reproduce the original paper's result & faster post-processing using …

…c++/swig

.gitignore

@@ -111,4 +111,6 @@ models/trained/*/checkpoint
 models/trained/*/*.pb
 models/trained/*/model-*.data-*
 models/trained/*/model-*.index
-models/trained/*/model-*.meta
+models/trained/*/model-*.meta
+models/pretrained/resnet_v2_101/eval.graph
+models/pretrained/resnet_v2_101/train.graph

README.md

@@ -29,6 +29,13 @@ You need dependencies below.
 - tensorflow 1.4.1+
 - opencv3, protobuf, python3-tk
 
+### Opensources
+
+- slim
+- slidingwindow
+  - https://github.com/adamrehn/slidingwindow
+  - I copied from the above git repo to modify few things.
+
 ### Install
 
 ```bash
@@ -143,6 +150,8 @@ See : [etcs/training.md](./etcs/training.md)
 
 [4] Keras Openpose : https://github.com/michalfaber/keras_Realtime_Multi-Person_Pose_Estimation
 
+[5] Keras Openpose2 : https://github.com/kevinlin311tw/keras-openpose-reproduce
+
 ### Lifting from the deep
 
 [1] Arxiv Paper : https://arxiv.org/abs/1701.00295

etcs/experiments.md

@@ -0,0 +1,9 @@
+
+
+## COCO Datasets
+
+| Model          | Scale      | AP         | AP 50      | AP 75      | AP medium  | AP large   | AR         | AR 50      | AR 75      | AR medium  | AR large   |
+|----------------|------------|------------|------------|------------|------------|------------|------------|------------|------------|------------|------------|
+| CMU            | Multi      |     0.5067 |     0.7660 |     0.5377 |     0.4927 |     0.5309 |     0.5614 |     0.7900 |     0.5903 |     0.5089 |     0.6347 |
+| CMU            | Single     |     0.5067 |     0.7660 |     0.5377 |     0.4927 |     0.5309 |     0.5614 |     0.7900 |     0.5903 |     0.5089 |     0.6347 |
+| Mobilenet thin | Single     |     0.2806 |     0.5577 |     0.2474 |     0.2802 |     0.2843 |     0.3214 |     0.5840 |     0.2997 |     0.2946 |     0.3587 |

models/pretrained/resnet_v2_101/download.sh

@@ -0,0 +1,5 @@
+#! /bin/bash
+
+wget http://download.tensorflow.org/models/resnet_v2_101_2017_04_14.tar.gz
+tar -xvf resnet_v2_101_2017_04_14.tar.gz
+rm resnet_v2_101_2017_04_14.tar.gz

requirements.txt

@@ -1,9 +1,10 @@
 argparse
 matplotlib
 scipy
+skimage
 tqdm
 requests
 fire
-ast
 dill
-git+https://github.com/ppwwyyxx/tensorpack.git
+git+https://github.com/ppwwyyxx/tensorpack.git
+slidingwindow

scripts/broadcaster_ros.py

@@ -16,9 +16,6 @@
 from networks import model_wh, get_graph_path
 
 
-scales = [None]
-
-
 def humans_to_msg(humans):
     persons = Persons()
 
@@ -52,8 +49,7 @@ def callback_image(data):
         return
 
     try:
-        global scales
-        humans = pose_estimator.inference(cv_image, scales)
+        humans = pose_estimator.inference(cv_image, resize_to_default=True, upsample_size=resize_out_ratio)
     finally:
         tf_lock.release()
 
@@ -63,31 +59,20 @@ def callback_image(data):
     msg.header = data.header
 
     pub_pose.publish(msg)
-    # rospy.loginfo(time.time() - et)
-
-
-def callback_scales(data):
-    global scales
-    scales = ast.literal_eval(data.data)
-    rospy.logdebug('[tf-pose-estimation] scale changed: ' + str(scales))
 
 
 if __name__ == '__main__':
-    global scales
-
     rospy.loginfo('initialization+')
     rospy.init_node('TfPoseEstimatorROS', anonymous=True, log_level=rospy.INFO)
 
     # parameters
     image_topic = rospy.get_param('~camera', '')
     model = rospy.get_param('~model', 'cmu')
+
     resolution = rospy.get_param('~resolution', '432x368')
-    scales_str = rospy.get_param('~scales', '[None]')
-    scales = ast.literal_eval(scales_str)
+    resize_out_ratio = float(rospy.get_param('~resize_out_ratio', '4.0'))
     tf_lock = Lock()
 
-    rospy.loginfo('[TfPoseEstimatorROS] scales(%d)=%s' % (len(scales), str(scales)))
-
     if not image_topic:
         rospy.logerr('Parameter \'camera\' is not provided.')
         sys.exit(-1)
@@ -106,7 +91,6 @@ def callback_scales(data):
     cv_bridge = CvBridge()
 
     rospy.Subscriber(image_topic, Image, callback_image, queue_size=1, buff_size=2**24)
-    rospy.Subscriber('~scales', String, callback_scales, queue_size=1)
     pub_pose = rospy.Publisher('~pose', Persons, queue_size=1)
 
     rospy.loginfo('start+')

src/common.py

@@ -100,17 +100,17 @@ def from_coco(human):
     (11, 12), (12, 13), (1, 0), (0, 14), (14, 16), (0, 15), (15, 17), (2, 16), (5, 17)
 ]   # = 19
 CocoPairsRender = CocoPairs[:-2]
-CocoPairsNetwork = [
-    (12, 13), (20, 21), (14, 15), (16, 17), (22, 23), (24, 25), (0, 1), (2, 3), (4, 5),
-    (6, 7), (8, 9), (10, 11), (28, 29), (30, 31), (34, 35), (32, 33), (36, 37), (18, 19), (26, 27)
- ]  # = 19
+# CocoPairsNetwork = [
+#     (12, 13), (20, 21), (14, 15), (16, 17), (22, 23), (24, 25), (0, 1), (2, 3), (4, 5),
+#     (6, 7), (8, 9), (10, 11), (28, 29), (30, 31), (34, 35), (32, 33), (36, 37), (18, 19), (26, 27)
+#  ]  # = 19
 
 CocoColors = [[255, 0, 0], [255, 85, 0], [255, 170, 0], [255, 255, 0], [170, 255, 0], [85, 255, 0], [0, 255, 0],
               [0, 255, 85], [0, 255, 170], [0, 255, 255], [0, 170, 255], [0, 85, 255], [0, 0, 255], [85, 0, 255],
               [170, 0, 255], [255, 0, 255], [255, 0, 170], [255, 0, 85]]
 
 
-def read_imgfile(path, width, height):
+def read_imgfile(path, width=None, height=None):
     val_image = cv2.imread(path, cv2.IMREAD_COLOR)
     if width is not None and height is not None:
         val_image = cv2.resize(val_image, (width, height))