[WIP] Machine Learning Classifier for NaviGator #414
Conversation
ghost
added
the
|
I think it is reasonable to target this for mil_common (save for the specific graphs/labelmaps for NaviGator's challenges). The goal is to have a node that implements the vision_node interface and takes in the necessary configuration files (graph, label map) as parameters. All this nodes has to do is implement a callback function which takes in an image and returns a list of labeled boxes. It is a class, so you can override init for startup configuration (loading in the graph). As for the large graph files, I recommend hosting them on plumbusi and using file(DOWNLOAD) in CMake to have it be downloaded on builds (verified with checksum) without having to have it as a git object. |
|
Another question: Is killing/restarting the processes the only solution for tensorflow's restrictions on multiple processes? For example, maybe you just need to close/restart the tensorflow session object? Worth a small amount of research to have a cleaner/faster solution. |
closes #456
closes #455
required by #360
required by #210
Depends on #138.
Very much a work in progress, this is basically a straight port from SubjuGator code. There are quite a few things that could improve this and cut down on file usage but as it stands this code functions and will return a classified image with bounding box.
TODO
Currently I have another branch that was previously integrated with PCODAR before its completion. I'm not sure how well it will work now but I have a baseline at least for the integration.