Woodscape Fisheye Motion Segmentation Challenge | CVPR 2023 OmniCV Workshop Forum

Go back to competition Back to thread list Post in this thread

> About the training images in evaluation

Hi,Can you give a more detailed description of the number of training images and validation scores?
I have some confusion about:
1. Does the number of synthetic images used in training lead to a penalty?
2. Is there a penalty for using more than 25% of WoodScapes images in training?
3. Is there a specific formula for calculating the verification score?

Thank ypu.

Posted by: heboyong @ April 15, 2023, 2:40 a.m.

Dear heboyong,

Yes, there is a penalty that is applied for using more than 25% of the Woodscape images for training or fine-tuning (Q2). Since for this challenge, we don't allow using pretrained models or training with other data, we encourage the contestants to train (and/or fine-tune) using Parallel Domain data and upto 25% of Woodscape data.

As for the exact formula for the score and for penalty regarding not using Parallel Domain data, I will get back to you as soon as possible (Q1 and Q3).

Regards,
Saravanabalagi.

Posted by: saravanabalagi @ April 17, 2023, 9:56 a.m.

Dear heboyong,

Thank you for your interest in our Parallel Domain & Valeo Motion Segmentation Challenge!

I have added brief information about the datasets in the Data section (https://codalab.lisn.upsaclay.fr/competitions/12131#participate-get_data).

Q1: No, only the number of WS images above 25% of the provided data.

Q3: The score depends on wIoU and a penalty associated with the usage of Woodscape images. Unfortunately we cannot reveal the exact formula. You can find the explanation here: https://codalab.lisn.upsaclay.fr/forums/12091/1662/

I hope the above answers your questions. If you have any further queries, please reply below.

Regards,
Saravanabalagi.

Posted by: saravanabalagi @ April 18, 2023, 5:39 p.m.
Post in this thread