[WIP] Tf object detection as grasp prediction

[ahundt]

I'm setting up to adapt the tf object detection API to find grasp bounding boxes. Here is what needs to happen:

• [x] install tensorflow/models repo with tf object detection api, see github.com/ahundt/robotics_setup
• [ ] try fine tuning pets dataset
• [x] generate tfrecord files for cornell dataset compatible with tf object detection api
• [ ] write script to download, configure and run training on cornell dataset
• [ ] run training, this will be limited because arbitrary rotation augmentation isn't possible
• [ ] visualize results and standard mAP scores
• [ ] if sufficiently good, integrate image augmentation with grasps into API
• [ ] if that also looks good, integrate and evaluate grasp_jaccard performance
• [ ] finally integrate into task apis
• [ ] ensure there are complete instructions to reproduce