tensorflow preferred[]
https://www.youtube.com/watch?v=MDP9FfsNx60 Tensorflow is preferred neural net.
install[]
https://github.com/facebookresearch/DetectAndTrack install into anaconda environment with centos6.6
https://caffe2.ai/docs/getting-started.html?platform=ubuntu&configuration=compile
http://caffe.berkeleyvision.org/installation.html and see errors at https://github.com/tonghe90/textspotter/issues/3. Recommends installing Nvidia cuda blade install with APT and not the runfile, else caffe breaks. Use gcc 5 or 4.9
notes[]
See http://caffe.berkeleyvision.org/ and its github https://github.com/BVLC/caffe Speed makes Caffe perfect for research experiments and industry deployment. Caffe can process over 60M images per day with a single NVIDIA K40 GPU* Object recognition with caffe connected to webcam . That’s 1 ms/image for inference and 4 ms/image for learning. We believe that Caffe is the fastest convolution net implementation available.
- https://www.youtube.com/watch?v=3sDpXoAXbeY Object recognition with caffe connected to webcam
upload images[]
http://demo.caffe.berkeleyvision.org/classify_upload classifies images you upload. Take snapshots from cctv stream on Infrared leds trigger classify as human, dog, cat ,car etc. It classifies animals very well, but for some reason it refuses to classify men and women, giving instead a generic "clothing" to "sports equipment". It can identify lipstick and spectacles on a women, but not the women seemingly.
embedded vision[]
- https://www.youtube.com/watch?v=XpRqqBmWHXk , http://www.embedded-vision.com/ Intel Graphics is pervasive on client platforms, and when well utilized, it can boost performance of a CNN-based application by factor of 5x compared with a default Caffe CPU + MKL implementation on the same device ("Cherry Trail" Atom x7-z8700, for instance). The demonstrated application is built on top of clCaffe (the Caffe framework extended with OpenCL), which is open sourced for easy reproducibility; its OpenCL code path is performance-optimized for Intel Graphics https://github.com/01org/caffe/tree/clcaffe
OpenKAI[]
- https://www.youtube.com/watch?v=BDgBvh3WLVw OpenKAI UAV platform running on JetsonTX1, using Caffe (SSD) for object detection https://github.com/yankailab/OpenKAI Open Kinetic AI: The base to implement vision and AI functions onto JetsonTX1 GPU as a companion to Pixhawk
- https://github.com/yankailab/OpenKAI install script and ubuntu package dependency.
- https://www.youtube.com/watch?v=5AVb2hA2EUs precision automatic landing of Quad frames,using openkai . See http://ardupilot.org/dev/docs/companion-computer-nvidia-tx1.html
- https://www.stereolabs.com/zed/specs/ 3d depth stereo camera perception to 20m.
Jetsonhacks[]
http://www.jetsonhacks.com/2017/03/02/daniel-tobias-car-cherry-autonomous-racecar/ and github cherry autonomous car
Intel github repo[]
https://github.com/01org , https://github.com/01org/caffe/tree/clcaffe caffe fork
install[]
https://gist.github.com/pjspillai/87625e1b6a129335eb363a4131a301f2
https://github.com/BVLC/caffe/blob/master/Makefile.config.example edit the makefile.config file . arch flag = 61 for gtx1060 and above. compile from source for applications that insist on a ROOT_DIR
http://caffe.berkeleyvision.org/installation.html