ROS Resources: Documentation | Support | Discussion Forum | Index | Service Status | ros @ Robotics Stack Exchange
Ask Your Question
0

What is the expected frequency of image objects in autoware.ai?

asked 2021-02-09 08:57:44 -0500

msmcconnell gravatar image

I am running the autoware.ai v1.13 (built in kinetic) base perception stack (lidar + camera with yolo weights). Currently I am building it with CUDA and running on an AStuff Spectra PC. The output frequency for the /detection/image_detector/objects topic seems to only be 3Hz. Is this expected? When using CUDA what is the expected frequency of this topic?

edit retag flag offensive close merge delete

1 Answer

Sort by » oldest newest most voted
0

answered 2021-02-10 13:57:31 -0500

Josh Whitley gravatar image

This depend on a lot of factors. You have to consider not only the upstream nodes/topics (the camera driver, the image rectifier, etc) but also the framerate that the camera itself is producing. So, in short, there is no "expected" frequency.

edit flag offensive delete link more

Comments

Expected might have been a poor choice of words. "Desired" might be more appropriate. I am wondering what the downstream components might desire to operate normally. For example my understanding is that most of the autoware.ai planning stack is designed to run at 10Hz. So I would think the object perception stack should operate at this same frequency. Does the fact this node is outputing at only 3Hz mean those nodes are in-capable of outputting data at the rate expected for planning or do they have logic to account for lapses in data on this topic?

msmcconnell gravatar image msmcconnell  ( 2021-02-10 14:36:32 -0500 )edit

The perception stack only operates at 10Hz because that is the typical frequency of data from a 3D lidar. If you had a lidar that ran faster, the downstream nodes would run faster as well. The design is almost completely data-driven (I say almost because there are some localization components which are temporal). The question you might want to ask instead ias "is the object detection node causing delays" which can be investigated by just looking at the frequency of the input and output topics. If the answer is no, look further upstream to see why the rate is so low.

If your camera is producing data at 30 Hz, I would excpet certain nodes like object recognition to be running slower but it's all up to hardware and what you consider safe.

Josh Whitley gravatar image Josh Whitley  ( 2021-02-10 14:44:12 -0500 )edit

Question Tools

1 follower

Stats

Asked: 2021-02-09 08:57:44 -0500

Seen: 165 times

Last updated: Feb 10 '21