Test 2 - Training Dataset Labels

attachment
training_GT_labels_v2.json
key insights
  • A JSON file containing the ground truth labels for the training dataset. You can find the Training Dataset here: https://www.herox.com/alphapilot/resource/314
  • File updated to v2 on 2/20/2019

comments

  • Zeryab Moussaoui Feb. 21, 2019, 2:20 p.m. PST
    Hello everyone,
    There are errors in some label order
    for example : IMG_8376 (1).JPG
    3rd and 4th corners are inversed.
    5% of images are like this one ..
  • Devin Feb. 19, 2019, 4:37 p.m. PST
    I also found that almost 1 out of 10 image is wrongly labeled, so be careful if you choose supervised learning model.
    • Kyla Jeffrey HeroX team Feb. 20, 2019, 5:37 p.m. PST
      Devin, the labels have been updated. Please see above.
    • Cory Pruce Feb. 20, 2019, 7:02 p.m. PST
      Hi Kyla, could we please have a list of which labels were updated?
    • Cory Pruce Feb. 20, 2019, 8:02 p.m. PST
      While the diff of the files is non-empty since the order has changed, my script yields no changes:

      import os
      import json

      prev_j = json.load(open('training_GT_labels.json'))
      cur_j = json.load(open('training_GT_labels_v2.json'))

      changed = {}

      for pj in prev_j:
      if pj not in cur_j:
      print('Missing', pj)
      prev = prev_j[pj]
      cur = cur_j[pj]

      if [(x,y) for x,y in zip(prev, cur) if x != y]:
      changed[pj] = cur

      with open('changed_labels.json', 'w') as outfile:
      json.dump(changed, outfile)
    • SnShines Feb. 21, 2019, 2:02 a.m. PST
      Yes, @Cory is right. Old file and this new file have the same ground truth values for all images.

      @Kyla, can you sort this issue and get us updated GT labels? Thanks.
  • Ed Cabrera Feb. 19, 2019, 11:36 a.m. PST
    Is 5199 mislabeled? "IMG_5199.JPG": [[]]
    • Daniel Barbosa Feb. 20, 2019, 8:38 p.m. PST
      I think that is correct because not all 4 corners are visible.
    • Taeyeon Kim Feb. 27, 2019, 1:17 a.m. PST
      This is a IMG_0562.JPG. Perhaps this is Gate 9 3/4.