darknet_ros
darknet_ros copied to clipboard
leggedrobotics
CPU real-time testing problem
Hello, i'm testing Yolov2-tiny only on CPU (also set GPU = 0), both on bag files and real time camera streams.
At first I cloned the leggedrobotics repo which resulted in asynchronized messages between the bounding boxes and detection images.
What I did to resolve this (as it was recommended on a different issue post) was to assign "boundingboxResults_.image_header.stamp" to the "boundingboxResults_.header.stamp" , which solved my problem only for bag files, so I had to find a solution for the real time asynchronization.
Issue posts (again) led me here. This version of YoloObjectDetector.cpp actually bypassed the asynchronization.
By modifying .cfg file, Im testing the detector on 320*320 network input resolution which results almost in ~1 FPS on Intel i7-8565u CPU @ 1,8GHz × 8.
Q1: Any rational explanation for bag files initial synchronization solution , not working for real time streams? Q2: Are my FPS values normal? Q3: Can FPS increase be achieved without modifying the network's input image resolution (which I know that works)? Q4: OpenMP = 1 or 0 inside /darknet_ros/darknet/Makefile seems to not affect the FPS. Is this really how Multithreading is enabled?
Sorry for long post!
Sorry for late reply
Q1: @mbjelonic might now Q2: On CPU that might be normal. Did you build in release mode? Q3: Build in release mode and try to tweak the source code to be more efficient Q4: I cannot confirm this at this point as most people are using their GPU, but you could validate by checking the CPU core utilization when running. If approx. 100% it's single threaded, if > 100% it's multithreaded.
Sorry for long post!
No excuse needed.
Q4: I think the "darknet_ros" didn't use the OPENMP default. You could insert the code into the file "darknet_ros/darknet_ros/CMakeLists.txt", if you want to use the openmp in your program.
find_package( OpenMP REQUIRED)
if(OPENMP_FOUND)
message("OPENMP FOUND")
set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}")
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}")
set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}")
endif()
Hello, i'm testing Yolov2-tiny only on CPU (also set GPU = 0), both on bag files and real time camera streams.
At first I cloned the leggedrobotics repo which resulted in asynchronized messages between the bounding boxes and detection images.
What I did to resolve this (as it was recommended on a different issue post) was to assign "boundingboxResults_.image_header.stamp" to the "boundingboxResults_.header.stamp" , which solved my problem only for bag files, so I had to find a solution for the real time asynchronization.
Issue posts (again) led me here. This version of YoloObjectDetector.cpp actually bypassed the asynchronization.
By modifying .cfg file, Im testing the detector on 320*320 network input resolution which results almost in ~1 FPS on Intel i7-8565u CPU @ 1,8GHz × 8.
Q1: Any rational explanation for bag files initial synchronization solution , not working for real time streams? Q2: Are my FPS values normal? Q3: Can FPS increase be achieved without modifying the network's input image resolution (which I know that works)? Q4: OpenMP = 1 or 0 inside /darknet_ros/darknet/Makefile seems to not affect the FPS. Is this really how Multithreading is enabled?
Sorry for long post!
Q4: I think the "darknet_ros" didn't use the OPENMP default. You could insert the code into the file "darknet_ros/darknet_ros/CMakeLists.txt", if you want to use the openmp in your program.
find_package( OpenMP REQUIRED) if(OPENMP_FOUND) message("OPENMP FOUND") set(CMAKE_C_FLAGS "${CMAKE_C_FLAGS} ${OpenMP_C_FLAGS}") set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} ${OpenMP_CXX_FLAGS}") set(CMAKE_EXE_LINKER_FLAGS "${CMAKE_EXE_LINKER_FLAGS} ${OpenMP_EXE_LINKER_FLAGS}") endif()
jap, this worked for me. i am running darknet_ros on CPU and I managed to increase the FPS from 1 to 2.5 by adding this line of code in darknet_ros/darknet_ros/CMakeLists.txt. Thanks a lot
