Added OBR-Autonomous Labels - Githubissues

ddavid / fsoco

Formula Student Objects in Context Dataset for the Formula Student Driverless competitions

Apache License 2.0

57 stars 53 forks source link

Added OBR-Autonomous Labels #55

Open BelleGarlick opened 4 years ago

BelleGarlick commented 4 years ago

Altered index.md, provided labels and logo. Cheers

ddavid commented 4 years ago

Hi Sam, am midst in the exam period, will go over it at the beginning of next week.

ddavid commented 4 years ago

Hi Sam, please excuse the delay, got a bit under the weather after the trip back to Munich. Will go through the PR later on tonight.

ddavid commented 4 years ago

Hi Sam, unfortunately there seems to be some labels without corresponding images. Could you please look into that? Maybe mismatched file names?

BelleGarlick commented 4 years ago

Sorry about that, have pushed a fix with the correct labels.

ddavid commented 4 years ago

Will probably only be able to have a look tomorrow, when I have access to WiFi again. Will keep you posted.

ddavid commented 4 years ago

Hi Sam, still having similar issues with labels in the repo not having a corresponding image in the album you shared with me. I've updated the draw_boxes script slightly. Feel free to use it to check if all labels and images are available. https://github.com/ddavid/fsoco/blob/master/scripts/draw_boxes_from_cwd.py Furthermore, there is a float class index in label 122.txt, haven't looked beyond that one yet but that would also need to be fixed.

BelleGarlick commented 4 years ago

Hi David.

I've updated the problem with class numbers incorrectly stores as floats. I'm not sure how you're having issues with mis-matched images/labels. I've re-uploaded all the images into the album (to ensure you have the correct images): https://photos.app.goo.gl/7ymWVc9L7m8nDttQ8. Additionally I've run the code you provided over the images which can be found here: https://drive.google.com/file/d/1R37sIMSoaTKgWsyPmCjGmHqUcMxgOLhu/view?usp=sharing. I hope that kelps.

ddavid commented 4 years ago

Hi Sam, I'll have a look tonight, thank you for going through it again.

ddavid commented 4 years ago

Hi Sam, thank you, the issue seems to be solved. I've looked into it on my side and I think it was some issue with mixed jpg and png images in one directory. The overall quality of the labels is really good, even more so through the HD images. Unfortunately there's still a few images with wrong labels and some where single cones were forgotten. Will send you a complete list here tomorrow morning.

ddavid commented 4 years ago

Hi Sam, I've gone through the first 300 images and have the feedback contained in the following list: 2020-03-10_partial-obu-feedback.txt All images in that range not mentioned are perfect as they are. Could you please revise the others and also apply similar changes to the remaining half I didn't go through yet? The recurring small issues were:

two cones labeled as one
far-away cones labeled imprecisely (boxes should be tight around the cone)
missing labels for cones (maybe some label files got clipped?)
very small portions of cones labeled (e.g. one edge of the base)

For the latter, approx. half of the cone should still be visible and the type of cone should be recognizable (harder for small vs. large orange cones) which is already the case in almost all of your images.