问题确认 Search before asking
请提出你的问题 Please ask your question
@laugh12321 Thank you for this nice and well working software repository! I'm using your software in my ROS2 Jazzy implementation, similarly as #160, but then in Python.
For my robotics application I need a segmentation mask for up to 50 objects per image, being able to process everything at video speed (<33 ms). When I use your TensorRT version scaled at image size 320, the model.predict time is 6 ms, which is extremely fast.
Unfortunately, the mask upscaling procedure as demonstrated in utils.py is the calculation bottleneck as my masks need to be upscaled to the original 720x1280 sized images (for up to 50 objects per image). Is there a faster method to do the post-processing after doing the TensorRT inference?
In addition, I was wondering whether the result.masks[i] output as in resized_mask could somehow only be the clipped mask or the flattened list clipped to the bounding box coordinates? That makes post-processing much simpler.
Thanks in advance!
问题确认 Search before asking
请提出你的问题 Please ask your question
@laugh12321 Thank you for this nice and well working software repository! I'm using your software in my ROS2 Jazzy implementation, similarly as #160, but then in Python.
For my robotics application I need a segmentation mask for up to 50 objects per image, being able to process everything at video speed (<33 ms). When I use your TensorRT version scaled at image size 320, the model.predict time is 6 ms, which is extremely fast.
Unfortunately, the mask upscaling procedure as demonstrated in utils.py is the calculation bottleneck as my masks need to be upscaled to the original 720x1280 sized images (for up to 50 objects per image). Is there a faster method to do the post-processing after doing the TensorRT inference?
In addition, I was wondering whether the result.masks[i] output as in resized_mask could somehow only be the clipped mask or the flattened list clipped to the bounding box coordinates? That makes post-processing much simpler.
Thanks in advance!