Working openpose python wrapper

Author: reinzor

image_recognition_openpose/CMakeLists.txt

@@ -1,72 +1,14 @@
 cmake_minimum_required(VERSION 2.8.3)
 project(image_recognition_openpose)
 
-## Add support for C++11, supported in ROS Kinetic and newer
-add_definitions(-std=c++11)
+find_package(catkin REQUIRED)
 
-find_package(catkin REQUIRED COMPONENTS
-  roscpp
-  cv_bridge
-  sensor_msgs
-  image_recognition_msgs
-  diagnostic_updater
-)
-
-catkin_package(
-  CATKIN_DEPENDS roscpp cv_bridge sensor_msgs image_recognition_msgs diagnostic_updater
-)
-
-set(OPENPOSE_DIR ${CMAKE_CURRENT_SOURCE_DIR}/openpose)
-IF(EXISTS ${OPENPOSE_DIR})
-  message("Using openpose from " ${OPENPOSE_DIR} " ...")
-
-  set(CAFFE_DIR ${OPENPOSE_DIR}/3rdparty/caffe)
-
-  set(CAFFE_INCLUDE_DIRS ${CAFFE_DIR}/distribute/include)
-  set(CAFFE_LINK_LIBRARY ${CAFFE_DIR}/distribute/lib/libcaffe.so)
-
-  set(OPENPOSE_INCLUDE_DIRS ${OPENPOSE_DIR}/include openpose/src /usr/local/cuda/include)
-  set(OPENPOSE_LINK_LIBRARY ${OPENPOSE_DIR}/build/lib/libopenpose.so)
-
-  add_definitions(-DUSE_CAFFE)
-
-  include_directories(
-    ${catkin_INCLUDE_DIRS}
-    ${CAFFE_INCLUDE_DIRS}
-    ${OPENPOSE_INCLUDE_DIRS}
-  )
+catkin_python_setup()
 
-  add_library(openpose_wrapper src/openpose_wrapper.cpp)
+catkin_package()
 
-  target_link_libraries(openpose_wrapper
-    ${CAFFE_LINK_LIBRARY}
-    ${OPENPOSE_LINK_LIBRARY}
-    ${catkin_LIBRARIES}
-  )
-
-  add_executable(${PROJECT_NAME}_test test/test.cpp)
-  target_link_libraries(${PROJECT_NAME}_test openpose_wrapper)
-else()
-  message(WARNING "Openpose folder not found in " ${OPENPOSE_DIR} " using a mock instead ...")
-
-  include_directories(${catkin_INCLUDE_DIRS})
-  add_library(openpose_wrapper src/openpose_wrapper_mock.cpp)
-
-  target_link_libraries(openpose_wrapper ${catkin_LIBRARIES})
-endif()
-
-# depend on generated messages
-add_dependencies(openpose_wrapper ${catkin_EXPORTED_TARGETS})
-
-add_executable(${PROJECT_NAME}_node src/openpose_ros_node.cpp)
-
-# Specify libraries to link a library or executable target against
-target_link_libraries(${PROJECT_NAME}_node openpose_wrapper)
-
-install(TARGETS
-  ${PROJECT_NAME}_node
-  openpose_wrapper
-  ARCHIVE DESTINATION ${CATKIN_PACKAGE_LIB_DESTINATION}
-  LIBRARY DESTINATION ${CATKIN_PACKAGE_LIB_DESTINATION}
-  RUNTIME DESTINATION ${CATKIN_PACKAGE_BIN_DESTINATION}
-)
+install(PROGRAMS
+  scripts/detect_poses
+  scripts/openpose_node
+  DESTINATION ${CATKIN_PACKAGE_BIN_DESTINATION}
+)

image_recognition_openpose/README.md

@@ -2,49 +2,85 @@
 
 ROS Wrapper for openpose https://github.com/CMU-Perceptual-Computing-Lab/openpose
 
-## Description
-Provides a service interface for openpose. Returns the skeleton when an image is send
+## Installation notes
 
-## Installation
-ROS Kinetic uses OpenCV 3.2 as default. Therefore it is important to compile openpose and caffe against OpenCV 3.2 as well
+This ROS wrapper makes use of the [Openpose python interface](https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/modules/python_module.md).
+Please follow the [installation manual](https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/installation.md) and ensure that the `BUILD_PYTHON` flag is turned on while running CMake.
 
-#### A simple example using the opencv 3.2 version distributed by ROS kinetic:
-```
-sudo apt remove opencv* libopencv*
-sudo apt install ros-kinetic-opencv3
-
-sudo ln -fs /opt/ros/kinetic/lib/libopencv_core3.so /usr/lib/libopencv_core.so
-sudo ln -fs /opt/ros/kinetic/lib/libopencv_highgui3.so /usr/lib/libopencv_highgui.so
-sudo ln -fs /opt/ros/kinetic/lib/libopencv_imgcodecs3.so /usr/lib/libopencv_imgcodecs.so
-sudo ln -fs /opt/ros/kinetic/lib/libopencv_imgproc3.so /usr/lib/libopencv_imgproc.so
-sudo ln -fs /opt/ros/kinetic/lib/libopencv_videoio3.so /usr/lib/libopencv_videoio.so
-sudo ln -fs /opt/ros/kinetic/lib/libopencv_objdetect3.so /usr/lib/libopencv_objdetect.so
-sudo ln -fs /opt/ros/kinetic/include/opencv-3.2.0-dev/opencv2 /usr/include/opencv2
-```
+## Scripts
 
-Next compile openpose using the [openpose installation manual](https://github.com/CMU-Perceptual-Computing-Lab/openpose/blob/master/doc/installation.md)
+### detect_poses
 
-Make sure at the end a symbolic link is added to the ROS package, for example if the openpose folder is in your home dir:
-```
-roscd image_recognition_openpose
-ln -s ~/openpose
+Example for the following picture:
+
+![Example](doc/example.jpg)
+
+```bash
+export MODEL_FOLDER=~/dev/openpose/models
+rosrun image_recognition_openpose detect_poses $MODEL_FOLDER image `rospack find image_recognition_openpose`/doc/example.jpg
 ```
 
-If the symbolic link is not present a mock node will be used for testing. 
+Output:
+
+![Example result](doc/example_result.jpg)
+
+It also works with a webcam stream, usage:
+
+```bash
+usage: detect_poses [-h] [--pose_model POSE_MODEL]
+                    [--net_input_size NET_INPUT_SIZE]
+                    [--net_output_size NET_OUTPUT_SIZE]
+                    [--num_scales NUM_SCALES] [--scale_gap SCALE_GAP]
+                    [--num_gpu_start NUM_GPU_START]
+                    [--overlay_alpha OVERLAY_ALPHA]
+                    [--python_path PYTHON_PATH]
+                    model_folder {image,cam} ...
+
+Detect poses in an image
+
+positional arguments:
+  model_folder          Path where the models are stored
+  {image,cam}           Mode
+    image               Use image mode
+    cam                 Use cam mode
+
+optional arguments:
+  -h, --help            show this help message and exit
+  --pose_model POSE_MODEL
+                        What pose model to use (default: BODY_25)
+  --net_input_size NET_INPUT_SIZE
+                        Net input size (default: -1x368)
+  --net_output_size NET_OUTPUT_SIZE
+                        Net output size (default: -1x-1)
+  --num_scales NUM_SCALES
+                        Num scales (default: 1)
+  --scale_gap SCALE_GAP
+                        Scale gap (default: 0.3)
+  --num_gpu_start NUM_GPU_START
+                        What GPU support (default: 0)
+  --overlay_alpha OVERLAY_ALPHA
+                        Overlay alpha for the output image (default: 0.6)
+  --python_path PYTHON_PATH
+                        Python path where Openpose is stored (default:
+                        /usr/local/python/)
+```
 
-(After creating the symlink, do not forget to clean first)
+### openpose_node
 
 ## How-to
 
 Run the image_recognition_openpose node in one terminal, e.g.:
 
-    rosrun image_recognition_openpose image_recognition_openpose_node _net_input_width:=368 _net_input_height:=368 _net_output_width:=368 _net_output_height:=368 _model_folder:=/home/ubuntu/openpose/models/
+```bash
+export MODEL_FOLDER=~/dev/openpose/models
+rosrun image_recognition_openpose openpose_node _model_folder:=$MODEL_FOLDER
+```
 
 Next step is starting the image_recognition_Rqt test gui (https://github.com/tue-robotics/image_recognition_rqt)
 
     rosrun image_recognition_rqt test_gui
-    
-Again configure the service you want to call with the gear-wheel in the top-right corner of the screen. If everything is set-up, draw a rectangle in the image and ask the service for detections:
+
+Configure the service you want to call with the gear-wheel in the top-right corner of the screen. If everything is set-up, draw a rectangle in the image and ask the service for detections:
 
 ![Test](doc/openpose.png)
 

image_recognition_openpose/doc/example.jpg

@@ -2,18 +2,17 @@
 <package format="2">
   <name>image_recognition_openpose</name>
   <version>0.0.0</version>
-  <description>The image_recognition_openpose package</description>
+  <description>ROS Wrapper for the openpose software package. Exposes a service and topic interface.</description>
 
   <maintainer email="[email protected]">Rein Appeldoorn</maintainer>
 
   <license>MIT</license>
 
   <buildtool_depend>catkin</buildtool_depend>
 
-  <depend>cv_bridge</depend>
-  <depend>diagnostic_updater</depend>
-  <depend>image_recognition_msgs</depend>
-  <depend>libopencv-dev</depend>
-  <depend>roscpp</depend>
-  <depend>sensor_msgs</depend>
+  <exec_depend>cv_bridge</exec_depend>
+  <exec_depend>diagnostic_updater</exec_depend>
+  <exec_depend>image_recognition_msgs</exec_depend>
+  <exec_depend>rospy</exec_depend>
+  <exec_depend>sensor_msgs</exec_depend>
 </package>

image_recognition_openpose/doc/example_result.jpg

@@ -0,0 +1,56 @@
+#!/usr/bin/env python
+
+import argparse
+import logging
+import sys
+
+import cv2
+
+from image_recognition_openpose import OpenposeWrapper
+
+parser = argparse.ArgumentParser(description='Detect poses in an image',
+                                 formatter_class=argparse.ArgumentDefaultsHelpFormatter)
+parser.add_argument('model_folder', help='Path where the models are stored')
+parser.add_argument('--pose_model', help='What pose model to use', default="BODY_25")
+parser.add_argument('--net_input_size', help='Net input size', default="-1x368")
+parser.add_argument('--net_output_size', help='Net output size', default="-1x-1")
+parser.add_argument('--num_scales', type=int, help='Num scales', default=1)
+parser.add_argument('--scale_gap', type=float, help='Scale gap', default=0.3)
+parser.add_argument('--num_gpu_start', type=int, help='What GPU support', default=0)
+parser.add_argument('--overlay_alpha', type=float, help='Overlay alpha for the output image', default=0.6)
+parser.add_argument('--python_path', help='Python path where Openpose is stored', default='/usr/local/python/')
+
+mode_parser = parser.add_subparsers(help='Mode')
+image_parser = mode_parser.add_parser('image', help='Use image mode')
+image_parser.set_defaults(mode='image')
+cam_parser = mode_parser.add_parser('cam', help='Use cam mode')
+cam_parser.set_defaults(mode='cam')
+
+# Image specific arguments
+image_parser.add_argument('image', help='Input image')
+
+args = parser.parse_args()
+
+logging.basicConfig(stream=sys.stdout, level=logging.DEBUG)
+
+wrapper = OpenposeWrapper(args.model_folder, args.pose_model, args.net_input_size, args.net_output_size,
+                          args.num_scales, args.scale_gap, args.num_gpu_start, args.overlay_alpha, args.python_path)
+
+if args.mode == 'image':
+    # Read the image
+    image = cv2.imread(args.image)
+    recognitions, overlayed_image = wrapper.detect_poses(image)
+
+    logging.info(recognitions)
+    cv2.imshow("overlayed_image", overlayed_image)
+
+    cv2.waitKey()
+elif args.mode == 'cam':
+    cap = cv2.VideoCapture(0)
+    while True:
+        ret, img = cap.read()
+        recognitions, overlayed_image = wrapper.detect_poses(img)
+        cv2.imshow("overlayed_image", overlayed_image)
+
+        if cv2.waitKey(1) & 0xFF == ord('q'):
+            break