[Major Feature] Training on surface relative transforms

[ahundt]

Train on surface relative transforms

Figure 6 in the Amazon Proposal pdf document illustrates surface relative transforms to gripper positions, when this task is complete we will be able to train using the grasp dataset's raw depth image and robot pose information.

Steps to complete this task:

• visualize point cloud in v-rep #293 visualize point cloud in v-rep, include a function to add points and to clear the points from a v-rep point cloud
• median filter with neighborhood rectangle parameter #292 create and configure a function to filter raw depth images so failed pixels (depth 0) are reduced or eliminated
• create a function def depth_image_pixel_to_cloud_point(depth_image, camera_intrinsics_matrix, pixel_coordinate, augmentation_rectangle=None) that can calculate xyz positions in the image frame when given a depth image and image intrinsic parameters
• Image Augmentation - random crop of depth and rgb image #291 Image Augmentation - Add support for random cropping of both an rgb and depth image, and integrate it into the training script.
• Surface relative transform calculation + visualization #294 Surface relative transform calculation + visualization - calculate the transforms for the grasping dataset. see grasp-viz releases for details.
• Grasp train: Image Resizing #360 Image resizing
• grasp_train.py integrate pixel-wise loss option #366 integrate pixel-wise loss into grasp_train.py
• Grasp Dataset - Fix ResNet model NaN loss #365 fix resnet model nan loss
• Add pixel-wise options to grasp model, Integrate Spatial Softmax #295 Add pixel-wise options to grasp model, Integrate Spatial Softmax
• 300 Epoch Training Run - Spatial Grasp Prediction #297 100 Epoch Training Run - Spatial Grasp Prediction

The Latest version of the V-REP scene file necessary for visualization is available in the fork of @ahundt. Code is currently maintained in the grasp_train_quaternion branch.

[ahundt]

Everything for this feature is done and merged, the pipeline runs!
Two outstanding bugs have been split out from this:
#402
#401

Closing