DIUx xView 2018 Detection Challenge
1. Object Classes
Here are the parent and child denominations for all 60 object classes in xView. Parent classes are at the headings of each column. The only exception is the last column 'None', which corresponds to classes that have no parent.
Fixed-Wing Aircraft |
---|
Small Aircraft |
Cargo Plane |
Passenger Vehicle |
---|
Small Car |
Bus |
Building |
---|
Hut/Tent |
Shed |
Aircraft Hangar |
Damaged Building |
Facility |
Truck |
---|
Pickup Truck |
Utility Truck |
Cargo Truck |
Truck w/Box |
Truck Tractor Trailer |
Truck w/Flatbed |
Truck w/Liquid |
Railway Vehicle |
---|
Passenger Car |
Cargo Car |
Flat Car |
Tank Car |
Locomotive |
Maritime Vessel |
---|
Motoboat |
Sailboat |
Tugboat |
Barge |
Fishing Vessel |
Ferry |
Yacht |
Container Ship |
Oil Tanker |
Engineering Vessel |
---|
Tower Crane |
Container Crane |
Reach Stacker |
Straddle Carrier |
Mobile Crane |
Dump Truck |
Haul Truck |
Scraper/Tractor |
Front Loader |
Excavator |
Cement Mixer |
Ground Grader |
Crane Truck |
None |
---|
Helipad |
Pylon |
Shipping Container |
Shipping Container Lot |
Storage Tank |
Vehicle Lot |
Construction Site |
Tower Structure |
Helicopter |
2. Examples
Example input, just image/pixel data:
Example output, with bounding boxes rendered visually:
3. Publication
Abstract
xView is a new large-scale dataset for the advancement of object detection techniques and overhead object detection research. This satellite imagery dataset enables research progress pertaining to four key computer vision frontiers. We utilize a novel process for geospatial category detection and bounding box annotation with three stages of quality control. Our data is collected from WorldView-3 satellites at 0.3m ground sample distance, providing higher resolution imagery than most public satellite imagery datasets. We compare xView to other object detection datasets in both natural and overhead imagery domains and then provide a baseline analysis using the Single Shot MultiBox Detector. xView is one of the largest and most diverse publicly available object detection datasets to date, with over 1 million objects across 60 classes in over 1,400 km2 of imagery.
The xVIEW dataset has been fully described in a paper which is freely available via arXiv.org: We encourage you to download and read the paper for more details.
Full satellite images gathered from DigitalGlobe’s WorldView 3 (WV3) satellites span a total area of over 1,400 square kilometers at 0.3 meter ground sample distance (GSD) resolution. The dataset was prepared in ways that are typical for satellite imagery, including ortho-rectification, pan-sharpening, RGB dynamic range adjustment, and atmospheric correction. xView contains more than 1 million labeled instances across 60 object categories. The fields in the geoJSON files include:
Field Name | Description |
---|---|
TYPE_ID | The bounding box label class ID |
CAT_ID | DigitalGlobe's unique ID for image strips |
IMAGE_ID | The image chip filename on which a feature is marked |
BOUNDS_IMCOORDS | Bounding box in pixel coordinates [xmin, ymin, xmax, ymax] of the image chip in which it is marked |
COORDINATES | Coordinates in longitude-latitude form for bounding box points |
The GeoTIFF files are broken into two sets: RGB and 8-band. The 8-band set contains 8-band multispectral GeoTIFF images labeled as image_id.tif where image_id is a unique integer. Similarly, the RGB set containes pan-sharpened 3-band RGB GeoTIFF images with the same naming convention.
The dataset was segmented into three sets, training, validation, and test, which have roughly 60, 20, and 20 percent of the images respectively. Restrictions were placed on the splits so that each set had enough instances of each category to perform a fair analysis. As with many challenges, all participants will have access to the train and validation images, but during the challenge, labels are provided for only the train dataset. The private test set will be used to evaluate final submissions after the Challenge deadline.
ZIP and TGZ files contain identical data, the only difference is the compression software used to create the archive. The size of correponding ZIP/TGZ files is approximately equal, choose the one that is more convenient for you. Please do not waste bandwidth downloading both.
Click the link (or "Save as...") to start downloading:
ZIP files
Download Training Images (zip)
SHA1: dce357ce3e7fff6ee28107e9211b793054979166
Download Training Labels (zip)
SHA1: 1897556cdc8cb24f7b137988c488768ed7d461b7
Download Validation Images (zip)
SHA1: 42c44878c4003c9411617ca05bf537aab41f420c
TGZ files (gzipped tarball)
Download Training Images (tgz)
SHA1: 1cdd5af68ec3a696f696104efb4730928cde524e
Download Training Labels (tgz)
SHA1: 693abfe34beabcb1bb87242241e0d0e273c36f0c
Download Validation Images (tgz)
SHA1: 1e030f1b2b35cdb8deb408a6d227a6713c32f7ba
Verify file integrity
SHA1 hashes are provided to verify integrity of your downloaded files. Compute the SHA1 checksum of your downloaded file using a command like shasum -a 1 FILENAME
, then compare the result computed from your local file to the SHA1 strings above. If they are identical, then your file is good.
ZIP decompression errors
If you encounter warnings or errors about offsets or "extra bytes at beginning of file" when decompressing these zip files, this is a known issue for decompressing files over 4 GB. The source files have been verified, but you may need to upgrade your decompression tool. On Windows, please use 7zip to decompress. On Mac, please use 7zip or The Unarchiver. On Linux you may need to upgrade your `unzip` binaries. Alternatively, download the gzipped tarball files (and decompress using a command like tar -xvzf FILENAME
).
Stale links and interrupted downloads
To maximize download speed, your download links are customized to stream data from a nearby server. If your download is interrupted, you may need to reload this page to obtain fresh download links. Please be respectful of download bandwidth, and re-download data only when necessary.