Touchdown Dataset

In collaboration with Jason Baldridge (Google), Eugene Ie (Google), Harsh Mehta (Google) and Yoav Artzi (Cornell University), we have enriched the StreetLearn dataset with 29k panoramas used in the Touchdown task and paper, covering a large rectangular area within Manhattan. Panoramas are available both as LevelDB archives of StreetLearn protobufs and as JPEG images.

The code to run the Touchdown task is available at https://github.com/lil-lab/touchdown. The details of the corpus and task are described in: Howard Chen, Alane Suhr, Dipendra Misra, Noah Snavely, and Yoav Artzi (2019), Touchdown: Natural Language Navigation and Spatial Reasoning in Visual Street Environments, CVPR.

Please refer to the dataset page to obtain both StreetLearn and Touchdown data. Please note that the natural language instructions are stored at https://github.com/lil-lab/touchdown, and not as part of our dataset, as they have been collected by Cornell University.

The following paper describes the integration of Touchdown into StreetLearn and its evaluation: Harsh Mehta, Yoav Artzi, Jason Baldridge, Eugene Ie, Piotr Mirowski (2020), Retouchdown: Adding Touchdown to StreetLearn as a Shareable Resource for Language Grounding Tasks in Street View, arXiv.