ROS Packages

- Estimator / Broadcaster
- Visualizer

Mobilenet_thin_463x368 graph added

CMakeLists.txt

@@ -24,7 +24,7 @@ generate_messages(
 )
 
 catkin_package(
-  CATKIN_DEPENDS message_generation message_runtime
+  CATKIN_DEPENDS rospy message_generation message_runtime
 )
 
-install()
+install()

etcs/training.md

@@ -96,4 +96,18 @@ $ bazel-bin/tensorflow/tools/graph_transforms/transform_graph \
     fold_constants(ignoreError=False)
     fold_old_batch_norms
     fold_batch_norms'
+```
+
+Also, It is promising to quantize neural network in 8 bit to get futher improvement for speed. In my case, this will make inference less accurate and take more time on Intel's CPUs.
+
+```
+$ bazel-bin/tensorflow/tools/graph_transforms/transform_graph \
+  --in_graph=/Users/ildoonet/repos/tf-openpose/tmp/cmu_640x480/graph_opt.pb \
+  --out_graph=/Users/ildoonet/repos/tf-openpose/tmp/cmu_640x480/graph_q.pb \
+  --inputs='image' \
+  --outputs='Openpose/concat_stage7:0' \
+  --transforms='add_default_attributes strip_unused_nodes(type=float, shape="1,360,640,3")
+    remove_nodes(op=Identity, op=CheckNumerics) fold_constants(ignore_errors=true)
+    fold_batch_norms fold_old_batch_norms quantize_weights quantize_nodes
+    strip_unused_nodes sort_by_execution_order'
 ```

launch/demo_video.launch

@@ -0,0 +1,44 @@
+<?xml version="1.0"?>
+<launch>
+    <arg name="video" default="$(find tfpose_ros)/etcs/dance.mp4" />
+    <arg name="video_width" default="1280" />
+    <arg name="video_height" default="720" />
+
+    <arg name="gpuid" default="0" />
+
+    <arg name="model" default="mobilenet_thin_432x368" />
+    <!--
+    <arg name="model" default="cmu_640x360" />
+    -->
+
+    <node name="pose_estimator" pkg="tfpose_ros" type="broadcaster_ros.py" output="screen" required="true">
+        <env name="CUDA_VISIBLE_DEVICES" value="$(arg gpuid)" />
+        <param name="camera" value="/videofile/image_raw" />
+        <param name="model" value="$(arg model)" />
+    </node>
+
+    <node name="pose_estimator_vis" pkg="tfpose_ros" type="visualization.py" output="screen" required="true">
+        <param name="camera" value="/videofile/image_raw" />
+        <param name="pose" value="/pose_estimator/pose" />
+        <param name="resize_ratio" value="0.3" />
+    </node>
+
+    <node name="image_view" pkg="image_view" type="image_view" respawn="false" output="screen" required="true">
+        <remap from="image" to="/pose_estimator_vis/output"/>
+        <param name="autosize" value="true" />
+    </node>
+
+    <include file="$(find video_stream_opencv)/launch/camera.launch">
+        <arg name="camera_name" value="videofile" />
+        <!-- means video device 0, /dev/video0 -->
+        <arg name="video_stream_provider" value="$(arg video)" />
+        <!-- throttling the querying of frames to -->
+        <arg name="fps" value="30" />
+        <!-- setting frame_id -->
+        <arg name="frame_id" value="videofile_frame" />
+        <arg name="width" value="$(arg video_width)"/>
+        <arg name="height" value="$(arg video_height)"/>
+        <arg name="visualize" value="true" />
+    </include>
+
+</launch>