Open Images Dataset v4 (Bounding Boxes)

A set of 1.9 million images, annotated with bounding boxes for 600 classes of objects, served in collaboration with Google.

Please view this page on a desktop computer to see more information including CSV preview and download options.

About this Data

NEW: We’re now hosting the Open Image Challenge 2018 Test set. You can download it by clicking this link.

Open Images is a dataset of 9 million images that have been annotated with image-level labels and bounding boxes spanning thousands of classes. Version 4 of Open Images focuses on object detection, with bounding boxes annotated across 1.9 million images. The annotations cover 600 classes of objects, grouped hierarchically.

We collaborated with Computer Vision scientists from Google to host this dataset, and it is the feature set for the “Open Images Challenge 2018”, the object detection challenge at the European Conference on Computer Vision 2018 (ECCV 2018).

For more on the Open Images Dataset, you can head here.




Annotated images from the Open Images dataset. Left: Mark Paul Gosselaar plays the guitar by Rhys A. Right: Civilization by Paul Downey. Both images used under CC BY 2.0 license.

Raw Data

The input data for this job is 9 million royalty-free images. It was created as part of a multi-job workflow. To recreate it, please contact Figure Eight.

000026e7ee790996test Itza Mexico7825604EEGQ0ZQ/OxJAIGzfJ6LMvw==
000062a39995e348test WilliamsParque Zoológico de São Paulo / Sao Paulo Zoo4227077oc+H0KgW0sxyEgAlfOiQDQ==
0000c64e1253d68ftest - >;<Redneck Figure 8's1744183Xufa/rhlOnBQSotR/8Gqhg==
000132c20b84269btest de Salud de Neuquenconsultorio ecografia piedra del Aguila (4)461828uPw/ycB1skL/byjF0AYJKA==
0002ab0af02e4a77test CowperKing Neptune scarecrow at La Seigneurie gardens on #Sark I looked a bit scarecrow-like myself after cycling round the island165352NRy8lYJlGUkHRRzeY65smw==
0002cc8afaf1b611test IngallsDemo224667MXSNPEBGjmYxSa8MG6z6Ww==
0003d84e0165d630test and SarahMissing flying saucer search5159064ts6/kaaAhoAOCWpUqL3W3w==
000411001ff7dd4ftest HarmanChicken friends on hike to fortress in Lopud1992202tIcFphdfWtmCcZcHDlup7A==
00045d609ca3f4ebtest X. O&#x27;NeilCaleb Carving Pumpkin6956054w7BUpcjmRDP96Z9XE+Z2Jg==
00068d5450f0358btest Jeffery BeggerlyIMG_2643602094RJukq9y/Vtb/25YxZPSuRA==
0006acf221a3e1d1test Mix @ Mi Va Me2077825JYvI721GQJfKb2hL/c3t+w==
000794c94c6d86eatest Quintanosorbet the venetian1664369VY9NoAWZpbGTeZ4D+lwQcg==
0007ef7b46ad6bb8test WardP10407576356992YgdG0PRRk8n5pBxqnynVuQ==
00090369f46e7b9etest TQ5 main board4925676dAFE0RuatBKjgp0Ttpq2Pw==
00094d5e8b3cb038test Dowsett vs Aman Ajula968796uCiGTF3Y57ztqVNOsNcOHw==
000a546e910f0a6btest from Balconey347895A/sdaKZt7nA+kPkArDJTKw==
000aa0b1c8fd5ddftest ChengHuddle4178888iUZaXZsMvVV1fGgQvHL99Q==
000b14e4ee4a2b2btest Guogmc_08sierra2500crewcab_angularfront_Large44061owo6r94uqqiLrDNDLsR8BA==
000b6e5bfa3e2a34test 343فطور العيد153932MmgTRYuP1uhG/z8jFUJleg==
000c5171b38d4bb0test;s5Tourist site4980736JASPMn1Ajjwb2K+IzdjOjA==
000c5f3f0b58ce18test IKEDA20150506_181631.jpg1348124TmZaK4PaX4k4sNRch5dnlA==
000cf5859025877ftest Mammy - there's a spider!11628074lb4JJlCqtJEGYGHtA53g==
000dbcc283d87f67test J.HayatThe Environment - تصوير عبدالعزيز جوهر حيات1362015AbaTmr9kbUV/AzJlvZtnRQ==
Download Options
NEW: Open Images Challenge 2018 Test Set - | 9.7 GB | 36.1 GB
test-annotations-bbox.csv | 49.8 MB
test-images.csv | 14.7 MB | 58.6 GB
train-annotations-bbox.csv | 1.1 GB
train-images-boxable.csv | 206.1 MB | 59.1 GB | 59.1 GB | 58.6 GB | 59.0 GB | 58.2 GB | 58.3 GB | 59.0 GB | 42.0 GB | 12.0 GB
validation-annotations-bbox.csv | 16.3 MB
validation-images.csv | 5.1 MB
Please note

The official Open Images Challenge 2018 Test Set is in file Do not confuse it with, which is the test set of Open Images V4.

The file class-description-boxable.csv should be used to identify class IDs and their corresponding object class names. You can download it here.

The train and validation sets contain the image and bounding-box annotations for Open Images V4. The training images also correspond to those used in the Open Images Challenge 2018, but not the box annotations. For the box annotations specific to the Challenge, visit the Challenge page.