ROS package for Detic. Run on both CPU and GPU, GPU is way performant, but work fine also with CPU (take few seconds to process single image).
example of custom vocabulary. Left: default (lvis), Right: custom ('bottle,shoe')
example of three dimensional pose recognition for cups, bottles, and bottle caps.
Ofcourse you can build this pacakge on your workspace and launch as normal ros package. But for those using CUDA, the following docker based approach might be safer and easy.
Prerequsite: You need to preinstall nvidia-container-toolkit beforehand. see (https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html)
Build docker image
git clone https://github.com/HiroIshida/detic_ros.git
cd detic_ros
docker build -t detic_ros .
Example for running node on pr1040 (please replace pr1040
by you robot hostname or localhost
):
python3 run_container.py -host pr1040 -mount ./launch -name sample.launch \
out_debug_img:=true \
out_debug_segimg:=false \
compressed:=false \
device:=auto \
input_image:=/kinect_head/rgb/image_color
The minimum necessary argument of run_container.py
is host
, mount
and name
:
out_debug_img:=true
). This launch args must come after the above three args.Another example for running three dimensional object pose detection using point cloud filtered by segmentation.
python3 run_container.py -host pr1040 -mount ./launch -name sample_detection.launch \
debug:=true \
vocabulary:=custom \
custom_vocabulary:=bottle,cup
Or rosrun detic_ros run_container.py
if you catkin build this package on the hosting computer side.
As in this example, by putting required sub-launch files inside the directory that will be mounted on, you can combine many node inside the container.
vocabulary:='custom' custom_vocabulary:='bottle,shoe'
. model_type
parameter.res50
), and avoid having too many classes in the frame (by e.g. setting a custom vocabulary or higher confidence thresholds). The sample_detection.launch
with default parameters handles all of this, yielding object bounding boxes at around 10Hz.Example for using the published topic from the node above is masked_image_publisher.py. This will be helpful for understanding how to apply SegmentationInfo
message to a image. The test file for this example also might be helpful.
See definition of srv/DeticSeg.srv
~input_image
(sensor_msgs/Image
)
~debug_image
(sensor_msgs/Image
)
~debug_segmentation_image
(sensor_msgs/Image
with 32SC1
encoding)
~segmentation_image
in grayscale image is almost completely dark and not good for debugging. Therefore this topic scale the value to [0 ~ 255] so that grayscale image is human-friendly.~segmentation_info
(detic_ros/SegmentationInfo
)
use_jsk_msgs
is false. Includes the class name list, confidence score list and segmentation image with 32SC1
encoding. The image is filled by 0 and positive integers indicating segmented object number. These indexes correspond to one plus those of class name list and confidence score list. For example, an image value of 2 corresponds to the second (index=1) item in the class name and score list. Note that the image value of 0 is always reserved for the 'background' instance.~segmentation
(sensor_msgs/Image
)
use_jsk_msgs
is true. Includes the segmentation image with 32SC1
encoding.~detected_classes
(jsk_recognition_msgs/LabelArray
)
use_jsk_msgs
is true. Includes the names and ids of the detected objects. In the same order as ~score
.~score
(jsk_recognition_msgs/VectorArray
)
use_jsk_msgs
is true. Includes the confidence score of the detected objects. In the same order as ~detected_classes
.As for rosparam, see node_cofig.py.
rosrun detic_ros batch_processor.py path/to/bagfile
See source code for the options.