Detic ros - A simple ROS wrapper for Detic instance segmentation using pre-trained dataset

This PR enables to specify vocabulary and custom_vocabulary from launch files.

• Sample usage

roslaunch detic_ros sample.launch out_debug_img:=true out_debug_segimg:=false compressed:=false device:=auto input_image:=/kinect_head/rgb/image_color vocabulary:='custom' custom_vocabulary:='bottle, shoe'

I think DetecWrapper should be just a function, and should not have any internal state. Based on this concept, the source code is refactored.

@Affonso-Gui

major

• [x] publish pcl_msgs/pointInices
• [x] publish/subscribe compressed image
• [ ] publish image bounding box
• [ ] fix up docker file (especially part where torch installation procedure)

minor

• [x] check installation instruction in README.md
• [x] use user-defined instance class names

The colors of the debug_image are inverted when the input image topic is not encoded in bgr8. I apologize for the lack of confirmation, but I suspect that the input to the Detic model is also inverted.

This problem can be solved by specifying the encoding of the CvBridge conversion. This is also done in jsk_recognition, which I used as a reference.

image topic not in bgr8 encoding

$ rostopic echo /hsrb/head_rgbd_sensor/rgb/image_rect_color --noarr
header: 
  seq: 25135
  stamp: 
    secs: 1668335381
    nsecs: 431184046
  frame_id: "head_rgbd_sensor_rgb_frame"
height: 480
width: 640
encoding: "rgb8"
is_bigendian: 0
step: 1920
data: "<array type: uint8, length: 921600>"
---

| before | after | | ---- | ---- | | | |

Comparison before and after change, the upper image is input image topic and the lower image is debug_image.

image topic in bgr8 encoding

$ rostopic echo /kinect_head/rgb/image_color --noarr
header: 
  seq: 39033
  stamp: 
    secs: 1668335714
    nsecs: 387932439
  frame_id: "head_mount_kinect_rgb_optical_frame"
height: 480
width: 640
encoding: "bgr8"
is_bigendian: 0
step: 1920
data: "<array type: uint8, length: 921600>"
---

| before | after | | ---- | ---- | | | |

Comparison before and after change, the upper image is debug_image and the lower image is input image topic.

The depth compression nodes here causes the error, and because of this entire launched nodes do not function https://github.com/HiroIshida/detic_ros/blob/e3596bd88c02f9f99f4dac2688741b6d886e2b57/launch/decompress_depth.launch#L6 [ERROR] [1668268511.903285160]: Unsupported image format: 32FC1; compressedDepth

Anyway, I fixed the error by building image_transport_plugins inside docker. But, I don't have time to find the cause of this.

This PR solve https://github.com/HiroIshida/detic_ros/issues/34

• Changed to install in the same way as other ros-noetic- packages.
• Avoid hang on keyboard country setting. https://stackoverflow.com/questions/63476497/docker-build-with-ubuntu-18-04-image-hangs-after-prompting-for-country-of-origin
• After consulting with @HiroIshida , I removed the comment out.

If we really want to comment in, do we need to change the wording in the README as well?

when trying to use launch/sample_detection.launch docker build failed when commenting out lines, https://github.com/HiroIshida/detic_ros/blob/2b2ccb94f37901d923b7c29f3afb9fefdf8173fb/Dockerfile#L49-L5 　 These are error messages

E: Unable to locate package ros-noetic-jsk-pcl-ros
E: Unable to locate package ros-noetic-jsk-pcl-ros-utils

All of log

• Update the detic submodule
• Add model_type option to select different real time models (https://github.com/facebookresearch/Detic/blob/main/docs/MODEL_ZOO.md#real-time-models)
• Add use_jsk_msgs option to publish the result as jsk_recognition_msgs types
• Add sample launch file of 3D object detection using the jsk_pcl_ros pipeline.
• Remove explicit background item from published topics for compatibility with jsk_recognition nodes.

With this I could get around 10Hz bounding box detection on the res50:

roslaunch detic_ros sample_detection.launch vocabulary:=custom custom_vocabulary:=cup,bottle,bottle_cap