Installation using newer CUDA versions & enforcing GPU mode

[tlind]

Hi and thanks for sharing the code! To save other people some time in reproducing the results:

In order to get the code running on newer versions of CUDA and cudnn (9.0 and 7.1.4.18 in my case), I applied the following pull request to the official caffe distribution, instead of using the 3-year old version that is linked as a submodule in rbgirshick's fast-rcnn: BVLC/caffe#4163. This pull request incorporates the necessary roi-pooling layers etc. and is able to load the provided models.

Initially, the detector was running extremely slow at only 0.2 Hz. I had to add a call to caffe.set_mode_gpu() immediately before the call to im_detect(...) in the message callback in detection_tracking.py. Now I reach around 15 Hz on a GTX 1080 Ti. The issue appears to be related to the threading model of rospy: rbgirshick/py-faster-rcnn#286

[marinaKollmitz]

Thank you very much for your helpful comments. I added the caffe.set_mode_gpu() call to detection_tracking.py in commit c285962. I also generated a fork of fast-rcnn which will be maintained for the hospital-people-detector. It does not include rbgirshick's caffe version any more. Instead, it links to BVLC's caffe with the pull request as you suggested or, as an alternative, our caffe fork which we will also maintain for the hospital-people-detector.