Darknet-ros process died on Jetson TX1

[pushkalkatara]

I am running Yolo v3 on Jetson TX1 but after loading the model in the memory and processing some frames from the Image topic the process dies.
I noticed in Tegra-stats, there is a constant increase in memory usage. I added a swap file of 16 GB also, but the memory fills up the swap with time and the process dies.
Some of the possible issues i thought of:

1. Memory leak between threads.
2. As the Yolo's output is around 3-5 fps and the camera's fps is 30, the image buffer is filling up and the ros publisher's queue is not rejecting frames.
3. Memory leak when subscribing to /videofile/image_raw topic ros-drivers/video_stream_opencv#15

The issue is somewhere related to ROS as the same model with the config file based on Yolo v3 was successfully running without ROS.

[pushkalkatara]

Small update related to the issue - https://answers.ros.org/question/220502/image-subscriber-lag-despite-queue-1/

[mbjelonic]

How did you solve this issues?

…

On 18 Oct 2018, at 15:27, Pushkal Katara ***@***.***> wrote: Small update related to the issue - https://answers.ros.org/question/220502/image-subscriber-lag-despite-queue-1/ — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[pushkalkatara]

I am still trying to figure out a way to solve the issue. As the implementation is using ImageTransport Subscriber in C++, there is no such thing as buff_size as an argument in roscpp.

[pushkalkatara]

Hi @mbjelonic, I tried to solve the issue using AlexAB darknet and different multithreading implementation. The issue is resolved for me. I have pushed the code on this link - https://github.com/pushkalkatara/darknet_ros.

[DeepDuke]

Except for "malloc" and "free" problems, I also met

[darknet_ros-1] process has died [pid 13109], exit code -11

I am quite confused about this problem. I am running this package on tx2 with a Intel Realsense Camera D415.