I haven't seen restrictions so far in the contest section. In particular, are the following prohibited:
-use of pretrained networks
-use of supplementary training images (outside of the test set)
-use of ensemble techniques
Thanks!Posted by: MasterOfTheForest @ May 12, 2021, 12:13 a.m.
thank you for asking using the forum for asking these questions.
It's important that this should be clarified before really approaching the competition:
1. Use of pretrained networks: This is definitely allowed, the TAS500 dataset on its own is probably too small to learn generalizable filters for your neural networks. But I would recommend adding the information on what dataset you pretrained your model weights in the 'description' field of your submission.
2. Use of supplementary training images (outside of the test set): Same as for question (1), you are allowed to use supplementary training data, but this should definitely be specified in your submission description. Here we would limit this to supplementary data, that is publicly available. Here it might be woth having a look at the datasets like RELLIS-3D, RUGD, and Yamaha-CMU Off-Road Dataset. I'll add links to these datasets on the competition page. You can also extend the training and validation set with additonal annotation information, but here we encourage making the altered version of the TAS500 dataset publicly available and to notify its existence in this forum.
3. Use of ensemble techniques: The use of ensemble techniques is allowed. If this increases the model inference time too much, then it might make your approach uninteresting for the more general research area of using this for outdoor robotics, but as part of this competition ensemble techniques are definitely allowed.
Thank you for asking in the forum!
I'll add this information onto the competition main page.
Don't hesitate to ask in case there are any remaining questions.