Dear Challenge Committee,
There are some training approaches that seem to be in the gray area for being a visual prior. I will try to give a few and ask your opinion whether these are allowed or not.
1. Suppose that I used the complete COCO object detection dataset (only annotations, not the images) and obtain some meta-statistics. Can we use this metadata in our training? This metadata acts as a set of visual priors that can be obtained from a multiple different sources, i.e. a knowledge bases, other vision datasets etc.
2. In the second scenario, I would like to use a pretrained vision model, apply that model to the challenge COCO images (~5800 of them) and create a processed image set. Then using this processed images, I train my network with the small dataset, in a pre-training stage to teach my model to learn some important aspect of vision (depth, segmentation etc.) Now with this approach, I have not used any pre-trained model, or checkpoint but utilized a trained model to enrich my small dataset by adding other non-rgb dimensions. So, I used the wisdom of other datasets and other models on my small COCO dataset. Is it considered as a prior?
In general, I think we need a more detailed definition of what a visual prior is. Maybe there is such definition somewhere, if so please share them because I am unaware of such.
Regards,
Ozgur
Hi Ozgur,
Thanks for your question! We understand the exact implementation of "visual priors" might be hard to nail down. Our intention with this challenge is to focus on solutions that implement a visual prior in the DNN/CNN itself, rather than in data. We feel the data hunger of DNNs is a fundamental problem and would like this challenge to focus on solving it in other ways than adding more data.
As for your specific questions, the answer to both questions is no. We only allow using the dataset we provide. Using the full COCO dataset is prohibited, as is using any other model-based pre-processing like you describe. Basically, as long as you only process the data provided by us, you are within the competition limits.
Best,
Robert-Jan