You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is.
We use the ZED ROS2 wrapper for handling a lot of the vision pipeline, but as of right now (though this may have changed or will change soon), the wrapper doesn't handle custom object detection.
Describe the solution you'd like
A clear and concise description of what you want to happen.
We need to write some nodes for interfacing with the output's of the ROS2 wrapper, and then for sending results back into it so it can integrate it into the wrapper's built in visual SLAM. We should try inter-process setup and intra-process (composable ROS2 nodes) to see which is faster. We will have to take images from the correct topics, run them through the object detector (ideally it is optimized using tensorrt), and then send the results back into the wrapper.
If the wrapper still doesn't have capabilities to handle custom object detections, we might have to change some of the wrapper codebase to implement it ourselves, or implement our own custom algorithms for 3D localization from the 2D detection results. This will have to be looked into.
Here is some task's to follow through with:
rebuild ROS2 wrapper to make sure everything is up to date
add a directory somewhere in perception folder where we store .pt and tensorrt model files
write mock data generator that just samples an image from a sample of the dataset and resizes it to 1280x720, then sends it on a topic
this is for testing the object detector's within ROS framework
will need to throw a sample of the dataset somewhere in the codebase as well for testing
write object detector node in python that reads images from mock data generator topics
just use .pt files and ultralytics for now, test inference results and speeds
optimize the models using tensorrt and compare the results with the .pt pipeline
start working on c++ implementation that interfaces directly with the wrapper code (this will be the most involved part probably)
Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.
Additional context
Add any other context or screenshots about the feature request here.
Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is.
We use the ZED ROS2 wrapper for handling a lot of the vision pipeline, but as of right now (though this may have changed or will change soon), the wrapper doesn't handle custom object detection.
Describe the solution you'd like
A clear and concise description of what you want to happen.
We need to write some nodes for interfacing with the output's of the ROS2 wrapper, and then for sending results back into it so it can integrate it into the wrapper's built in visual SLAM. We should try inter-process setup and intra-process (composable ROS2 nodes) to see which is faster. We will have to take images from the correct topics, run them through the object detector (ideally it is optimized using tensorrt), and then send the results back into the wrapper.
If the wrapper still doesn't have capabilities to handle custom object detections, we might have to change some of the wrapper codebase to implement it ourselves, or implement our own custom algorithms for 3D localization from the 2D detection results. This will have to be looked into.
Here is some task's to follow through with:
Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.
Additional context
Add any other context or screenshots about the feature request here.
Some potentially helpful links:
The text was updated successfully, but these errors were encountered: