Improve the Overall FPS with Bandwidth Optimization

[visionsense-support]

The current bottleneck in the system is caused by publishing high-resolution images to ROS topics. Due to network bandwidth limits, this introduces latency. To address the bandwidth issue, the disparity node is merged with the stereo camera node and the ROS message publishing frequencies and queuing limits should be adjusted to help improve the frame rate.

[visionsense-support]

To optimize the ROS communication and reduce latency, change the following publisher frequencies as suggested:

1. node_camera_stereo.cpp: publisher queue size=10 -->2
2. node_stereo_depth.cpp: publisher queue size=10 -->2
3. node_gui.cpp: publisher queue sizes= 5 -->2, 10-->2
4. node_detect.cpp= 10-->2
5. node_lanedet.cpp=10-->2

Using these changes you will observe improvement, but you can play around further. Increase the queue size for some nodes will actually improve the overall FPS but in others keeping the value minimum is better.

[garrethklauer]

Could Possibly changing the detection models also help with latency.

[pirazor]

@garrethklauer, I don’t think your unit is anywhere near full GPU utilization yet; it should comfortably handle larger models.

Also, don’t read too much into the latency you’re seeing in the visualization node. Each perception node is publishing asynchronously with very low latency. The small delay comes from the final visualization step, which waits for all streams to resynchronize. That node is intended for system debugging, not for showcasing end-to-end perception performance.

In most AV stacks, teams run their own asynchronous HMI/visualization and subscribe directly to individual perception outputs. We’re following the same approach and have a separate 3D graphical interface in development.

If you can share more about your end goal, we can give more targeted guidance. For example, if you’re focused on object and lane detection only, a single synchronous node would give you near-zero latency with the current global-shutter camera.