Can we use external dataset to pretrain our model? I noticed that it is said we can only use the given dataset (its RGB/RGB-D information), so I wonder if we can use external datasets to pretrain our model and use optical flow information to improve the recognition accuracy.Posted by: wuyongfa @ Dec. 30, 2020, 2:09 p.m.
the only restriction we impose is that RGB track must use RGB data only, as well as RGB+D track needs to deal with RGB+D data only. In addition to this, you can use any feature or additional data in pretraining. Note, in the case you use additional data, you will need to provide the data or your pretrained model at the code verification stage, so that the organizers can re-train the model if needed, in order to reproduce the results.
May I ask why you choose to show the latest result on the leaderboard instead of the best one?Posted by: wuyongfa @ Jan. 4, 2021, 1:09 p.m.
thanks for your question.
The reason for this choice is to give real-time feedback to participants to every single submission, otherwise, they may not see the result if the submission didn't improve a previous one.
It is their responsibility to monitor their submissions and upload before the deadline the one they consider will give the best score.
I hope this answers your question.
Hi, I wanted to ask whether we can use RGB-D data in the RGB track to pre-train some parts of our recognition system. But at testing the system will not need the RGB-D data anymore. Two hypothetical examples:
1) We use RGB-D data to train a system that will estimate depth from RGB. Then at the testing phase, we will first estimate depth from RGB and then recognize.
2) We use RGB-D data to estimate the 3D pose of a skeleton, then use these 3D poses to eg. cluster RGB data and use this clustering for training a sub-system (eg. to discriminate between poses). Hence, RGB-D was only used to pre-process data.
Thanks!Posted by: hpoes @ Feb. 3, 2021, 10:18 a.m.
RGB+D data is not allowed in any format and stage of training in RGB track.