Custom Model

Pretrained TrailNet (Redtail)

To use a custom model, such as the one used in the Redtail project, simply use post_proc:=None. In this case the node will just publish the raw output of the DNN (e.g. softmax layer) using standard Image messages.

• for 1D output like that of the softmax layer, the dimensions of the published image are 1x1xC where C is the size of the layer (e.g. 6 for the TrailNet out layer or 1000 for ImageNet's softmax layer).
• The encoding field of the message is in the following format: 32FCXXX where XXX is equal to number of channels. For example, for TrailNet: 32FC6, for ImageNet: 32FC1000. 32 means 32-bit and F means float. See the ROS documentation for more details.
• The data field contains the output of the DNN in byte array representation so it should be cast to a 32-bit float type.

The default topic name is /caffe_ros/network/output but it can be changed using launch file or command line.

The node requires two parameters:

• prototxt_path - path to the Caffe model .prototxt file
• model_path - path to the Caffe model binary file Additionally, make sure that other parameters, like the input/output layer names (input_layer/output_layer) are set correctly. An example of the command line for TrailNet:

rosrun caffe_ros caffe_ros_node __name:=trails_dnn _prototxt_path:=/data/src/redtail/models/pretrained/TrailNet_SResNet-18.prototxt _model_path:=/data/src/redtail/models/pretrained/TrailNet_SResNet-18.caffemodel _output_layer:=out

Note: when using rosrun remember to prefix parameters with _ (__ for system parameters) and use := when assigning a value.

To check that the DNN node is producing some data, make sure that camera node (a real camera via gscam or an image_pub node) is running, and execute rostopic echo /trails_dnn/network/output. An example of the output:

header: 
  seq: 201
  stamp: 
    secs: 1501106944
    nsecs: 796543294
  frame_id: ''
height: 1
width: 1
encoding: 32FC6
is_bigendian: 0
step: 24
data: [110, 74, 230, 58, 135, 125, 95, 60, 228, 14, 124, 63, 133, 26, 6, 61, 156, 75, 115, 63, 91, 87, 138, 60]

To convert the data byte array to float array, use this simple Python script:

import struct
data = bytearray([110, 74, 230, 58, 135, 125, 95, 60, 228, 14, 124, 63, 133, 26, 6, 61, 156, 75, 115, 63, 91, 87, 138, 60])
print(struct.unpack('<%df' % (len(data) / 4), data))