jhu-lcsr/costar_plan

Failures misclassified as success in data collection!

ahundt opened this issue · 3 comments

TODO:

  • Write Success/failure re-labeling script, use either
    • view_convert_dataset.py or
    • vrep_costar_stack.py. Some successful stacks are labeled as failures, need to write a script to fix it.
  • For each file show last image in sequence and press 1 to rename success or 2 to rename failure.

Some of the failures in this run are actually successes, we need to show them in a viewer to confirm, reprocess it, and correct the label:

2018-04-25-18-59-19_example000001.failure.h5f
2018-04-25-19-01-15_example000002.failure.h5f
2018-04-26-15-02-53_example000001.failure.h5f
2018-04-26-15-04-14_example000002.failure.h5f
2018-04-26-15-05-45_example000003.failure.h5f
2018-04-26-15-07-20_example000004.success.h5f
2018-04-26-15-08-42_example000005.failure.h5f
2018-04-26-15-10-04_example000006.success.h5f
2018-04-26-15-11-36_example000007.failure.h5f
2018-04-26-15-12-56_example000008.success.h5f
2018-04-26-15-15-16_example000009.failure.h5f
2018-04-26-15-16-40_example000010.failure.h5f
2018-04-26-15-18-02_example000011.failure.h5f
2018-04-26-15-19-23_example000012.failure.h5f
2018-04-26-15-20-45_example000013.failure.h5f
2018-04-26-15-22-11_example000014.failure.h5f

This is an example of a success misclassified as a failure:

python scripts/view_convert_dataset.py --preview True --path ~/.costar/data/2018-04-26-15-22-11_example000014.failure.h5f

I wrote a small script to try and correct the problems but it seems the recorded block poses aren't directly aligned with the base link frame so I can't directly use XYZ offsets for comparison, I'll need to plot the transforms and look at the issue in more detail to fix it.

Done as of 3f6dbfb in grasp_pixel_wise_training branch, plus much more!