shoichi-hasegawa0628 / detic_ros Goto Github PK

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):

docker run --rm --net=host -it --gpus 1 detic_ros:latest \
    /bin/bash -i -c \
    'source ~/.bashrc; \
    rossetip; rossetmaster pr1040; \
    roslaunch detic_ros sample.launch \
    out_debug_img:=true \
    out_debug_segimg:=false \
    compressed:=false \
    device:=auto \
    input_image:=/kinect_head/rgb/image_color'

Change the pr1040 part and /kinect_head/rgb/image_color in command above by your custom host name and an image topic. If compressed image (e.g. /kinect_head/rgb/image_color/compressed) corresponding to the specified input_image is also published, by setting compressed:=true, you can reduce the topic pub-sub latency. device is set to auto by default. But you can specify either from cpu or cuda.

Example for running three dimensional object pose detection on pr1040 :

docker run --rm --net=host -it --gpus 1 detic_ros:latest \
    /bin/bash -i -c \
    'source ~/.bashrc; \
    rossetip; rossetmaster pr1040; \
    roslaunch detic_ros sample_detection.launch \
    debug:=true \
    vocabulary:=custom \
    custom_vocabulary:=bottle,cup'

Add additional arguments to the script above. example: vocabulary:='custom' custom_vocabulary:='bottle,shoe'.

Detic is trained in different model types. In this repository you can try out all of the real-time models using the model_type parameter.

For higher recognition frequencies turn off all debug info, run on GPU, decompress topics locally, use smaller models (e.g. 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)
  • Input image
• ~debug_image (sensor_msgs/Image)
  • debug image
• ~debug_segmentation_image (sensor_msgs/Image with 32SC1 encoding)
  • Say detected class number is 14, ~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)
  • Published when 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)
  • Published when use_jsk_msgs is true. Includes the segmentation image with 32SC1 encoding.
• ~detected_classes (jsk_recognition_msgs/LabelArray)
  • Published when 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)
  • Published when 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.