Fix https://github.com/jsk-ros-pkg/jsk_recognition/issues/2860

[mqcmd196]

close #2860

libsiftfast has so many static global variables, so it is not thread-safe when called from multiple nodelets. This PR adds a global mutex to restrict the access to those variables to make it thread-safe.

Since the entire mutex is taken, if n imagesift nodelet are loaded, the time required to compute each feature is n times longer. But I think it is fine because the node is implemented with ConnectionBasedNodelet. If the user doesn't subscribe, the slow calculation doesn't occur. Making the user code work is much important.

The clean method is to make libsiftfast thread-safe, but I think we should write it from scratch or use OpenCV if we do so.

You can try the bug occurs just reverting 365b2d0 and execute test_two_imagesift.test

[mqcmd196]

requests.exceptions.ConnectionError: HTTPConnectionPool(host='jenkins.jsk.imi.i.u-tokyo.ac.jp', port=8080): Max retries exceeded with url: /crumbIssuer/api/json (Caused by NewConnectionError('<urllib3.connection.HTTPConnection object at 0x7f22e150c5d0>: Failed to establish a new connection: [Errno -5] No address associated with hostname',))

[mqcmd196]

CI green

[k-okada]

can you compare calculation speed in single thread with and whitout this PR? because libsiftfast is originally developed to pursue high speed processing by several optimization technique, so I'd like to know how this changes the performance.

[mqcmd196]

calculation speed in single thread

What do you mean? Making the nodelet's num_worker_thread=1 or making the libsiftfast's OpenMP flag OFF?