Open taeil opened 3 years ago
I downloaded the manifest file and found that the folder structure is similar to Eurostat. Images are within their class-specific folders. Then downloaded some JSON files from within the image folders. There are no geo-coordinates in the JSON files. Found that they were intentionally removed as part of the challenge 'Certain fields, such as GPS coordinates, were removed from all sets during the challenge......... now that the challenge has ended, the sequestered and ground truth data has been released, which contains all raw metadata, including category labels and GPS coordinates, for every image...'. It looks like the missing information is in the torrent file. Folder structure:
I found groundtruth.tar.bz2 file and extracted it which has train, val, test_gt and seq_gt with additional information. It has the following additional information: "raw_location": "POLYGON ((12.274391273000001 45.845910428099998, 12.2788828494 45.845910428099998, 12.2788828494 45.849049360099997, 12.274391273000001 45.849049360099997, 12.274391273000001 45.845910428099998))", "spatial_reference": "GCS_WGS_1984", "epsg": "4326",
The first one in raw_location is longitude and not latitude. The format is raw_location": "POLYGON ((longitude latitude...). Otherwise, it doesn't work.
Did some basic validation, using coordinates from fMoW groundtruth json file, corresponding image in the fMoW, and the map based on the coordinates. They match!
Created a tool which parses the fMoW dataset and creates a csv file with image name, its latitude, longitude, timestamp and country code(not required, but added).
Note: . This is the main file that is required if we want to generate the TFRecord or extract an image from different time from google earth engine API
Using fmow-rgb dataset instead of fmow-full. fmow-full is 3.5 tb. Will use fmow-full based on the need.
Ur master file should work same as full file. :)
Correct! It will work.
input: Seems like fMoW data use geo-coordinate.
output
The dataset is on s3. Instructions are on github.