What is this?

Classify object image from input image and mask using regional feature outputted by ResNet.

../../_images/regional_feature_based_object_recognition.gif ../../_images/regional_feature_based_object_recognition_objects.jpg

Subscribing Topic

  • ~input (sensor_msgs/Image)

    Input label image.

  • ~input/mask (sensor_msgs/Image)

    Region of interest.

Publishing Topic

  • ~output (jsk_recognition_msgs/ClassificationResult)

    Classification result of input image.


  • ~db_file (String, required)

    DB file which has the pairs of object label and ResNet feature vector.

  • ~gpu (Int, default: 0)

    GPU id to be used.


The sample classifies 39 objects which is used Amazon Picking Challenge 2016.

roslaunch jsk_perception sample_regional_feature_based_object_recognition.launch  # CPU mode
roslaunch jsk_perception sample_regional_feature_based_object_recognition.launch gpu:=0  # GPU mode

How to create db_file?

You can create the DB file form pairs of object image and mask for each object you’d like to recognize. In the sample, the db_file is automatically downloaded, but you can try to create it again in your environment.

rosrun jsk_perception \
  $(rospack find jsk_perception)/sample/data/apc2016_object_imgs_and_masks_templates \
  $(rospack find jsk_perception)/sample/data/resnet_features_apc2016.npz