Baselines and analysis: LVIS arXiv update

[Ross Girshick]

Hi folks,

We have two announcements.

1) Evaluation server. Thanks for your patience while we set up the evaluation server for the challenge. We will be using EvalAI and it should be available later this week. We've had some delays while a new feature to support large results files is implemented. When it's online, you will be able to submit val set results and test set results. You will be able to see your val results (they will not be public) so that you can debug the process of generating and submitting results. Test set results will not be visible to anyone other than the organizers until the challenge workshop.

2) Baselines and analysis. Today we updated the LVIS paper on arXiv to include two new appendix sections that will be of great interest to anyone who's participating in the challenge. Appendix B includes two Mask R-CNN baselines on the LVIS val set and Appendix C includes an analysis of how results transfer from the val set to the test set that will be used in the challenge. Please find the updates in v2 here: https://arxiv.org/pdf/1908.03195.pdf.

Best regards,

Ross, Agrim, and Piotr

[张松阳]

Hi, Ross

I would like to know the baseline in the updated paper use class-agnostic regression in the rcnn head or the class-specific regression head? 

Thanks.

[Ross Girshick]

The baselines use class-specific box regression.