grasp dataset bug in image cropping and resizing

[ahundt]

When images are being cropped + resized the projected gripper points in the vrep visualization are incorrect. This means the offsets aren't being handled properly for those cases.

Here is a screenshot demonstrating the bug:

Expected behavior is that the two black line segments for each pair of spheres should be collinear, piercing the gripper ball and ending in the surface ball.

Here is a screenshot of the correct behavior, when random_crop is False.
(Side note: this picture is from a different grasp attempt than the photo above)

To see the correct behavior, simply set the random_crop value to False. True will display the bug.
Here is the location:
https://github.com/cpaxton/costar_plan/blob/8864d838154675595a651ac7eb2b7e49630cad13/costar_google_brainrobotdata/grasp_dataset.py#L79

There are two tasks for this issue:

• correct the bug when random_crop == True
• correct any issues occurring when resize == True

[ahundt]

@DingYu95 are you interested in working on this issue?

[ahundt]

This bug is fixed:

However, code to extract pixel wise training tensors must still be implemented.