Submission Guidelines

Participants should submit a single image for each left frame in the test datasets. The naming of the image files should exactly match the source image. Each pixel of the image should be labelled with the class prediction using the color value in the training set, these value are also provided in the labels.json file. 

Submission FAQs

  • Can I use any other training data? Pre-trained networks are permitted, but only publicly available training data may be used. 
  • Can I submit more than one method? If you have two different techniques which you would like to submit, please send a request email to the organizers asking for permission. In general this will be approved however more than two different techniques will not be approved. 
  • How will submissions be evaluted? Submissions will be evaluated using the mean intersection-over-union criteria for all classes that are present in a given frame. The average over all frames will be computed to obtain an overall score. 
  • Will the evaluation scripts and test data be made available after the challenge? No, however submissions after the challenge will be processed on request. 
  • What is the penalty if I accidentally fail to upload a submission for 1 frame? We will try to alert you to upload the missing frame. However, if this is not successful you will be penalized by getting an error for that frame which is 2x the largest error of  the other team's method scores on that frame.

To make a submission, please use the following upload link. Please archive your submission images into a folder with your name (or team name) in the filename. You must have registered for the challenge and include the signed agreement to the subchallenge rules as part of your archive in the upload.

If you have issues using the upload function or need to edit your submission, please email