Jason Baldridge Zack Berger Yonatan Bitton Jaemin Cho Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang DOCCI is available on Hugging Face Datasets!
huggingface.co/datasets/googl…
Jason Baldridge This is a joint work with Sunayana Rane, Zack Berger, Yonatan Bitton, Jaemin Cho, Roopal Garg, Alexander Ku, zarana parekh, Jordi Pont-Tuset, Garrett Tanzer, Su Wang, and Jason Baldridge!
Check out our dataset in our project website: google.github.io/docci/
Yasumasa Onoe Zack Berger Yonatan Bitton Jaemin Cho Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang Jason Baldridge Vishaal Udandarao seems like this dataset could be useful for your research!
Emiel van Miltenburg Jason Baldridge Jing Yu Koh Zack Berger Yonatan Bitton Jaemin Cho @ ICLR2024🇦🇹 Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang Interesting point! We are not planning to release it at the moment as it requires some verification work, but we will consider it!
Emiel van Miltenburg Jason Baldridge Jing Yu Koh @ ICLR 🇦🇹 Zack Berger Yonatan Bitton Jaemin Cho @ ICLR2024🇦🇹 Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang Thanks for your interest! In short, Appendix A describes image curation. In Appendix B, we've listed key visual features that we tried to include in the descriptions.
Yasumasa Onoe Zack Berger Yonatan Bitton Jaemin Cho Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang Jason Baldridge Congrats on the release Yasu, amazing work! Glad to see that Jason Baldridge's crazy photos finally see the light of day 😛
Jason Baldridge Jing Yu Koh Yasumasa Onoe Zack Berger Yonatan Bitton Jaemin Cho @ ICLR2024🇦🇹 Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang The dataset looks super cool though! I’d be interested to study the patterns of inclusion: what information do annotators select to put into the descriptions?
Yasumasa Onoe Jason Baldridge Jing Yu Koh Zack Berger Yonatan Bitton Jaemin Cho @ ICLR2024🇦🇹 Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang Small note: I don't understand this part: 'Who was involved in the data collection process and how were they compensated? We disclose this upon acceptance.' -- this seems odd since it touches upon validity/reliability of the dataset. Shouldn't that info be available to reviewers?
Jason Baldridge Jing Yu Koh Yasumasa Onoe Zack Berger Yonatan Bitton Jaemin Cho Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang I’m wondering to what extent this collection is reproducible. The documentation does not seem to be detailed enough to instruct others to create a similar set of photographs.
Would be nice to have more detailed instructions for others to contribute their own photos!
Jing Yu Koh Yasumasa Onoe Zack Berger Yonatan Bitton Jaemin Cho Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang Yes, it's been a long while coming!! See Appendix A for my discussion/admission of obsession/craziness. :-)
Jason Baldridge Jing Yu Koh @ ICLR 🇦🇹 Yasumasa Onoe Zack Berger Yonatan Bitton Jaemin Cho @ ICLR2024🇦🇹 Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang Hmm.. it seems not all annotations are available. Would be nice to have stage 1,2&3 annotations (without further processing) to understand the description process better.
Also I cannot find the “detailed annotation guidelines” anywhere?
Yasumasa Onoe Jason Baldridge Jing Yu Koh Zack Berger Yonatan Bitton Jaemin Cho @ ICLR2024🇦🇹 Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang Thanks! I’m aware of the information you wanted to include in the descriptions. I’d like to study what actually ended up in the descriptions.
Could you make the info for all three stages available instead of only the final annotations? (Apologies if I missed them in the README.)
Yasumasa Onoe Jason Baldridge Jing Yu Koh Zack Berger Yonatan Bitton Jaemin Cho @ ICLR2024🇦🇹 Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang Thanks! Let me know when you do release the data; I think it might be a fun resource for NLG people since one could model the process of going from facts (Stage 1) to initial descriptions (Stage 2). Very similar to WebNLG, for example, except here there may be content selection.
Yasumasa Onoe Jason Baldridge Jing Yu Koh @ ICLR 🇦🇹 Zack Berger Yonatan Bitton Jaemin Cho @ ICLR2024🇦🇹 Roopal Garg Alexander Ku zarana parekh Jordi Pont-Tuset Garrett Tanzer Su Wang To be clear: I meant the raw annotations from all three stages, so I'm a bit confused by the verification requirement.