For context, here is @yoavartzi's tweet last year about the original Touchdown paper and dataset. It's a really hard and important pair of problems.https://twitter.com/yoavartzi/status/1075838661328035840 …
-
-
Prikaži ovu nit -
And here is
@MirowskiPiotr's tweet about releasing the initial version of StreetLearn. Our extension with Touchdown maintains the same access process and licensing.https://twitter.com/MirowskiPiotr/status/1102668151630761984 …
Prikaži ovu nit -
Even though both StreetLearn and Touchdown covered parts of Manhattan, only a small number of the panoramas were common in both datasets. We put the 29k panoramas referenced in Touchdown through manual vetting for privacy to ensure faces and license plates are blurred.pic.twitter.com/QDcVGKTcbT
Prikaži ovu nit -
This not only brings the number of panoramas in StreetLearn from 114k to 144k, it also provides some new possibilities for testing generalization from one set of panoramas in NYC to another.
Prikaži ovu nit -
Importantly, it allows Google to manage take down requests for panoramas and update researchers of any changes. We encourage the research community to use only vetted and approved resources like StreetLearn for their Street View oriented work.
Prikaži ovu nit -
Finally, we provide open source implementations for Touchdown vision-and-language navigation and spatial description resolution, as part of our VALAN code base. https://github.com/google-research/valan/tree/master/touchdown … (Complete instructions for reproducing the results are coming soon!)
Prikaži ovu nit -
This is a collaboration between Google Research (
@n0royalroad, Eugene Ie, me), DeepMind (@MirowskiPiotr) and Cornell Tech (@yoavartzi). We look forward to seeing how the research community builds on the data and produces new solutions for both the StreetLearn and Touchdown tasks!Prikaži ovu nit
Kraj razgovora
Novi razgovor -
Čini se da učitavanje traje već neko vrijeme.
Twitter je možda preopterećen ili ima kratkotrajnih poteškoća u radu. Pokušajte ponovno ili potražite dodatne informacije u odjeljku Status Twittera.
, a natural language navigation and spatial reasoning dataset using Street View. The task: follow the instructions to reach a goal and find a hidden