image
imagewidth (px)
50
640
creator_username
stringlengths
2
142
hash
stringlengths
25
32
gps_latitude
float32
-82.95
76.1
gps_longitude
float32
-176.28
179
date_taken
unknown
Ivan+Mlinaric
51cf7f6697426c264f8b1f3327446765
null
null
"2010-09-09T22:46:08"
Ryan+Wenner
afc3c77546e622e6d6fb3d9863acb2c
null
null
"2009-10-05T15:03:50"
Katatoniq
5ee216574fbb1aa147bcc41dbe99c9e4
null
null
"2010-08-26T12:37:12"
ingridtaylar
288570a378cba96cc3b1252c3d476e97
null
null
"2008-10-31T22:19:31"
Fran%C3%A7ois+Terrier
681ad5e9e6625d35da2af04438da3d
null
null
"2011-11-14T10:18:25"
cepoverde
bf64993239c67b2ab39ef466d12037a9
41.846
-6.86044
"2008-09-14T16:52:34"
DieselDemon
697c35b27667fa69bf3591cf7595095
34.395729
-119.71365
"2009-03-09T18:57:13"
Andrew.Beebe
f3f48abfd1d73d43eaaba36f389bf5
45.503639
-73.560397
"2009-08-08T14:44:34"
Solarbotics
1dcd55f143d25a9042af50829a1b479
null
null
"2012-11-09T13:50:42"
ldljuancho
8232361e4bd8f312d7fc35f24c473ec
null
null
"2005-12-18T11:02:48"
DSebourn
3ed8c054fdabcc16e94972fcd53981ee
null
null
"2011-02-11T14:59:50"
Eran+Finkle
1b682833bfff72049da60b08437c3b
null
null
"2011-03-26T23:58:20"
Colin+Godbout
3dffe385d2b082c310e8d36ffcb65f3
null
null
"2011-10-03T07:44:13"
jthoel
ec61623ef30e1bb4139804c75a6623d
38.968952
-77.40632
"2011-01-27T04:05:28"
UglyPlantation
24dcb04fce77d167105c76b5fd9daac
null
null
"2011-04-15T10:18:41"
MaxMalkov
188b743d132bbddb9dbd45fb31c270c5
null
null
"2012-11-20T14:21:31"
rweait-osm
ded1ee72333ca6c77679cae6df6147a
null
null
"2009-08-18T18:56:12"
matt+hutchinson
63b3c39a822ccc6dae16df13e42f13
53.363463
-6.280537
"2010-07-25T14:23:00"
Cimm
3bbde0646526f1a358f9827954feb78
50.850093
4.623803
"2007-05-23T20:52:43"
whispr03
c2cf9eb5ca8b4ed65eb552bff63dcd
null
null
"2010-10-23T13:42:15"
Fallonious+Monk
db6cab36dd499a68cfac5c7d66a7f7db
null
null
"2012-01-02T11:42:20"
dvrosa
28f75ea69385d56dc731fd1e1d42d68
null
null
"2012-06-14T09:18:48"
Denis+Bocquet
2dd9fed2e0a861c58be20ca253dbb7
48.918156
2.350816
"2011-11-28T22:32:01"
jasonbkk
177ee94ec44feb6e663fcec1388546f9
1.183955
124.568417
"2013-10-26T11:57:58"
ffg
dee399a6b45ace6849ab68a318055d
null
null
"2006-11-14T01:24:51"
Bobbie
bb8179caa983287e6bf92e03c893eaa
null
null
"2008-05-04T16:45:03"
_MaK_
47bd76b41c6d322c7d2f5493fa8321
23.700572
90.388705
"2011-11-18T16:28:30"
chrisbuckridge
c4d11b75c8273256432680e1533ca1a4
50.089588
14.417254
"2007-11-08T15:31:00"
Jeremy+Buckingham+MLC
ec4562fd38cce5281b8ee8bb8443f12d
null
null
"2012-07-16T05:24:01"
merkwelt
e8946a037df87a61356ad6c201f4f3d
null
null
"2009-02-19T22:53:45"
Youharman
75f9d9fb6a39b6999b92a822682b6ec
null
null
"2012-05-25T14:58:21"
Sodaro+Estate+Winery
ff8b8a741cdb5b48198fcd1b599237f
null
null
"2008-09-11T19:04:45"
rykerstribe
9e679953c7c6921f1cbbfa80ce5db3
47.809018
9.644386
"2008-03-15T18:42:46"
jam_232
487b7a4d3b387d68c741d4782e899245
null
null
"2012-11-16T11:11:39"
mary+churchill
634b994280b2ab195551bd5ffa76bc
42.324951
-71.080341
"2010-12-30T22:01:43"
Pedro+Angelini
d63a341538f7b0fdd5f5ac5c1a9430c7
-23.645153
-46.62117
"2008-08-24T11:35:19"
J.Feinberg+Photography
c9c6f621a73e6167d36dceb831bce8e4
40.766914
-73.974176
"2011-07-09T00:12:54"
omi_black
9016ed34cc9fc170fc92b0dd8487220
43.690125
13.265497
"2010-07-08T20:14:39"
ePi.Longo
15754e92a4dc6896df03e6ef7f3285
10.366256
107.072753
"2008-07-26T08:22:20"
amirygh
1a5c45e8c8d147b895d27996ee34ff4
null
null
"2012-01-24T05:05:05"
lucacorsato
c761a30802b81ed8322f9130f97e
null
null
"2012-11-26T11:00:52"
Trams+aux+fils.
4bcd263461a1e731143be3242f9fe0c9
46.316287
6.964044
"2010-10-22T11:50:52"
Saad.Akhtar
523c2e48b16e27e75d6277c21929a7f6
null
null
"2005-07-07T16:05:28"
Imagens+%C3%93rf%C3%A3s
9d2ffef5c9265578af5f276a742b9c1
-23.024447
-44.338073
"2014-04-24T21:09:59"
jhannon1
a3817589125c4e56b6d4a6b3057cc8
null
null
"2013-04-07T16:37:22"
DaMongMan
14c12964d6a2adfe7bb8ff24b9ad30
46.275175
-79.430851
"2010-08-03T21:25:17"
Spider+Joe
5662da96522790c06b4a115993e56717
null
null
"2007-09-20T08:39:26"
thegarethwiscombe
ca5bbac4dbb16ece4ab701f5e219091
null
null
"2006-10-23T23:09:08"
pattrick.muller
48f57ff312569b7c4f9b7c2e36f2345
null
null
"2010-10-17T11:29:28"
Marcos+Bola%C3%B1os
cb7a84c65559fbe5c7372ef7fcaf81d
null
null
"2013-11-02T21:11:34"
JBCLINTON
89abd0d969e44b1bec74f4cc98c9c17
null
null
"2014-03-10T18:40:10"
DrummondBruce
5a788ecaf74e5c4be2ac9441628e0cb
29.958039
-90.063536
"2009-08-21T13:08:18"
brokentrinkets
64e3bcf46653a6b2ba20afd4fbab861c
43.550833
7.010833
"2011-08-01T13:38:46"
gudmd.haralds
9a1e53257243ea4546d559182ced8312
null
null
"2008-08-04T15:48:53"
the_girl
e7659ed86e87c5335ec7b70af10f74d
null
null
"2005-07-07T20:24:33"
Eric+Grossnickle
4b88acc0c7482e3cb1b82660865e2
null
null
"2009-07-22T22:59:59"
MicGloWal
90ae226e74c7415f232a7bb947b88d
-34.830009
20.00988
"2007-01-08T12:36:38"
picdrops
089c930569940ea2158c07f5e5150d1
null
null
"2007-12-28T00:06:41"
comedy_nose
5b23c6f52cc986af9de97da4b6597d
null
null
"2014-01-19T11:42:58"
lee_kmp
a141f79f1652b8ad64505a4fdffc30b0
null
null
"2009-08-02T10:26:27"
manojdarolia
ceb8acbfc92fab73245e781588dfab
null
null
"2010-12-27T00:37:34"
raitank
994ffb422b74bc6919bdc517ec468368
null
null
"2008-10-06T12:29:33"
Wheezer+Geezer
668c8a6de995e234c55a028db985f
null
null
"2012-08-14T00:00:01"
bardgabbard
de3658bf78938285ccfb1a29876d6a1
null
null
"2009-10-10T07:19:18"
Philheckel
b319394db3d144f3df1391f3432598
43.558127
4.060821
"2012-09-06T07:11:37"
Rui+Ornelas
ee369f23d9f62e810259978387754f9
39.020988
-9.178175
"2007-05-26T15:14:37"
Paul+Davidson
46489377aad82d1f6e1e7fb66ebda16
null
null
"2005-11-24T16:33:36"
ChairWomanMay
f6d8ac77d0c486557d9d2c46b8539792
null
null
"2011-04-24T15:41:42"
Aiko%2C+Thomas+%26+Juliette%2BIsaac
cf7be1b1e5a66a431a799fec97af7a36
null
null
"2013-08-08T13:50:31"
Juan+Lupi%C3%B3n
da22169e4ab611f51955e2e5bc5687b
null
null
"2006-08-24T11:49:46"
gabofr
bfd3bd64fec1de80a1145a8c3412c79
32.559148
-116.435873
"2009-06-20T16:48:51"
Thomas+Ormston
8588fe5a02f7a58f9e78249d3d98b4
49.888152
8.648127
"2013-09-08T13:10:02"
failinth
86c4fa75d986fb6c1edd1bca523e09c
null
null
"2010-12-07T01:46:56"
standardpixel
e4e384a2b6becf929e41328747e3589d
37.810614
-122.261888
"2010-10-16T13:15:02"
craigdfreeman
4cae6e85a3f6f8d8a0f27b193588e46e
51.523371
-0.114348
"2014-01-09T11:26:02"
theogeo
fbaa4a69a8d7fc3e63db993c8c60c8b7
null
null
"2007-01-15T14:13:32"
Curimedia+%7C+P+H+O+T+O+G+R+A+P+H+Y
6adac869859036c7137c6ec2d389383
41.310396
2.108216
"2012-08-17T17:30:31"
pverlegh
27ec33485c63dfa565fb6b3595d7758
null
null
"2009-07-23T10:10:45"
bamvivid
8e32163083c9e7e7222014c287cc5b
43.025928
141.398817
"2012-02-14T22:57:42"
Erik+Kruszewski
6da291a3c7f0bc7dbf51aeaa5a25262
40.863874
-73.876919
"2013-08-18T11:31:58"
Fake+Plastic+Alice
645bf1ae72bfb6c5d3eb46e879c0f2
null
null
"2008-11-04T07:45:23"
T_Monk
bef76bd8a07448c167924ceca4bc7d7
null
null
"2011-04-08T01:14:09"
Vimages
51db18ad58f970a58f8dc9ca8562dbe5
null
null
"2008-12-29T15:46:18"
frugal+dattier
9b4ff14189667c262805d84de3623df
null
null
"2013-02-27T05:50:46"
roman.petruniak
2a826badb48f1917fd46678d77dfc6ac
null
null
"2009-01-10T22:36:31"
nledez
531a8a81febee3c4fb3fcfb5318db8
46.103155
3.885802
"2009-08-08T17:15:41"
geezaweezer
405a95e854fd96e26d7e8d4f61c5859
null
null
"2009-08-24T17:09:16"
erwinsoo
57552375ac8c6d1d7f1b16aca228ee
1.285979
103.859038
"2012-11-17T10:40:56"
nateOne
8edc058e5917f2e83edafd1c2bf59
null
null
"2008-09-19T18:20:20"
Holding+Steady
aacf9580ee5d2b586c18b622116d3128
null
null
"2010-10-02T18:48:38"
CesarKuriyama
14b218452d6d9769d74b4b9e8ec598c
null
null
"2006-05-01T00:00:00"
Chez+Eskay
da61c988f5481c7609d241feee4d9db
54.819827
-3.837919
"2008-10-02T17:23:28"
Napafloma-Photographe
e214285e2cd3ba481426181c6910e6e9
50.297566
2.7759
"2014-02-07T16:04:18"
Canadian+Starhawk
a8b65b347b98516a3243f1d265daea6
null
null
"2010-05-27T15:19:06"
mikedarnell1974
1739977d73a2546e75858f9d64926bd
32.891768
35.409107
"2011-03-22T16:37:33"
lukar14
fce831e5159166d6f532d7a5114cf7
null
null
"2010-02-06T06:54:41"
tortxof
adf9ea3cd5b0b154ec494c903d781568
null
null
"2007-06-29T08:52:49"
turbo.beagle
e2771705e74d71be4bad5d19e3de3
null
null
"2012-06-01T20:38:16"
Lisa+J+Jackson
513082cc844a6570b826ed8b2e71183b
null
null
"2013-09-15T02:23:06"
Rubbermaid+Products
cbf721a5632cd896c51d027f34ecf4b
null
null
"2012-03-13T12:06:56"

Dataset Card for PASS

Dataset Summary

PASS is a large-scale image dataset, containing 1.4 million images, that does not include any humans and which can be used for high-quality pretraining while significantly reducing privacy concerns.

Supported Tasks and Leaderboards

From the paper:

Has the dataset been used for any tasks already? In the paper we show and benchmark the intended use of this dataset as a pretraining dataset. For this the dataset is used an unlabelled image collection on which visual features are learned and then transferred to downstream tasks. We show that with this dataset it is possible to learn competitive visual features, without any humans in the pretraining dataset and with complete license information.

Is there a repository that links to any or all papers or systems that use the dataset? We will be listing these at the repository.

What (other) tasks could the dataset be used for? We believe this dataset might allow researchers and practitioners to further evaluate the differences that pretraining datasets can have on the learned features. Furthermore, since the meta-data is available for the images, it is possible to investigate the effect of image resolution on self-supervised learning methods, a domain largely underresearched thus far, as the current de-facto standard, ImageNet, only comes in one size.

Is there anything about the composition of the dataset or the way it was collected and preprocessed/cleaned/labeled that might impact future uses? Given that this dataset is a subset of a dataset that randomly samples images from flickr, the image distribution is biased towards European and American creators. As in the main papers discussion, this can lead to non-generalizeable features, or even biased features as the images taken in other countries might be more likely to further reflect and propagate stereotypes [84], though in our case these do not refer to sterotypes about humans.

Are there tasks for which the dataset should not be used? This dataset is meant for research purposes only. The dataset should also not be used for, e.g. connecting images and usernames, as this might risk de-anonymising the dataset in the long term. The usernames are solely provided for attribution.

Languages

English.

Dataset Structure

Data Instances

A data point comprises an image and its meta-data:

{
  'image': <PIL.JpegImagePlugin.JpegImageFile image mode=RGB size=500x375 at 0x7FFAD48E35F8>, 'creator_username': 'NTShieldsy',
  'hash': 'e1662344ffa8c231d198c367c692cc',
  'gps_latitude': 21.206675,
  'gps_longitude': 39.166558,
  'date_taken': datetime.datetime(2012, 8, 9, 18, 0, 20)
}

Data Fields

  • image: A PIL.Image.Image object containing the image. Note that when accessing the image column: dataset[0]["image"] the image file is automatically decoded. Decoding of a large number of image files might take a significant amount of time. Thus it is important to first query the sample index before the "image" column, i.e. dataset[0]["image"] should always be preferred over dataset["image"][0].
  • creator_username: The photographer.
  • hash: The hash, as computed from YFCC-100M.
  • gps_latitude: Latitude of image if existent, otherwise None.
  • gps_longitude: Longitude of image if existent, otherwise None.
  • date_taken: Datetime of image if existent, otherwise None.

Data Splits

All the data is contained in the training set. The training set has 1,439,588 instances as this implementation corresponds to the most recent release (v3) from the version history.

From the paper:

Are there recommended data splits (e.g., training, development/validation, testing)? As outlined in the intended usecases, this dataset is meant for pretraining representations. As such, the models derived from training on this dataset need to be evaluated on different datasets, so called down-stream tasks. Thus the recommended split is to use all samples for training.

Dataset Creation

Curation Rationale

From the paper:

For what purpose was the dataset created? Neural networks pretrained on large image collections have been shown to transfer well to other visual tasks where there is little labelled data, i.e. transferring a model works better than starting with a randomly initialized network every time for a new task, as many visual features can be repurposed. This dataset has as its goal to provide a safer large-scale dataset for such pretraining of visual features. In particular, this dataset does not contain any humans or human parts and does not contain any labels. The first point is important, as the current standard for pretraining, ImageNet and its face-blurred version only provide pseudo-anonymity and furthermore do not provide correct licences to the creators. The second point is relevant as pretraining is moving towards the self-supervised paradigm, where labels are not required. Yet most methods are developed on the highly curated ImageNet dataset, yielding potentially non-generalizeable research.

Source Data

Initial Data Collection and Normalization

From the paper:

  • Collection process:

    How was the data associated with each instance acquired? The data was collected from the publicly available dataset YFCC-100M which is hosted on the AWS public datasets platform. We have used the meta-data, namely the copyright information to filter only images with the CC-BY licence and have downloaded these using the aws command line interface, allowing for quick and stable downloading. In addition, all files were subsequently scanned for viruses using Sophos SAVScan virus detection utility, v.5.74.0.

    What mechanisms or procedures were used to collect the data (e.g., hardware apparatus or sensor, manual human curation, software program, software API)? Our dataset is a subset of the YFCC-100M dataset. The YFCC-100M dataset itself was created by effectively randomly selecting publicly available images from flickr, resulting in approximately 98M images.

    Does the dataset contain all possible instances or is it a sample (not necessarily random) of instances from a larger set? The dataset is a sample of a larger set—all possible digital photographs. As outlined in Section 3 we start from an existing dataset, YFCC-100M, and stratify the images (removing images with people and personal information, removing images with harmful content, removing images with unsuitable licenses, each user contributes at most 80 images to the dataset). This leaves 1.6M images, out of which we take a random sample of 1.28M images to replicate the size of the ImageNet dataset. While this dataset can thus be extended, this is the set that we have verified to not contain humans, human parts and disturbing content.

    Over what timeframe was the data collected? The images underlying the dataset were downloaded between March and June 2021 from the AWS public datasets’ S3 bucket, following the download code provided in the repo. However the images contained were originally and taken anywhere from 2000 to 2015, with the majority being shot between 2010-2014.

  • Preprocessing/cleaning/labeling:

    Was any preprocessing/cleaning/labeling of the data done (e.g., discretization or bucketing,tokenization, part-of-speech tagging, SIFT feature extraction, removal of instances, processing of missing values)? After the download of approx. 17M images, the corrupted, or single-color images were removed from the dataset prior to the generation of the dataset(s) used in the paper. The images were not further preprocessed or edited.

    Was the “raw” data saved in addition to the preprocessed/cleaned/labeled data (e.g., to support unanticipated future uses)? Yes. The creators of the dataset maintain a copy of the 17M original images with the CC-BY licence of YFCC100M that sits at the start of our dataset creation pipeline. Is the software used to preprocess/clean/label the instances available? We have only used basic Python primitives for this. For the annotations we have used VIA [27, 28].

Who are the source language producers?

From the paper:

Who was involved in the data collection process (e.g., students, crowdworkers, contractors) and how were they compensated (e.g., how much were crowdworkers paid)? As described, the data was collected automatically by simply downloading images from a publicly hosted S3 bucket. The human verification was done using a professional data annotation company that pays 150% of the local minimum wage.

Annotations

Annotation process

This dataset doesn't contain annotations.

Who are the annotators?

This dataset doesn't contain annotations.

Personal and Sensitive Information

From the paper:

Does the dataset contain data that might be considered confidential (e.g., data that is protected by legal privilege or by doctor-patient confidentiality, data that includes the content of individuals’ non-public communications)? No.

Does the dataset contain data that, if viewed directly, might be offensive, insulting, threatening, or might otherwise cause anxiety? No. Besides checking for human presence in the images, the annotators were also given the choice of flagging images for disturbing content, which once flagged was removed.

Does the dataset relate to people? If not, you may skip the remaining questions in this section. No.

Does the dataset identify any subpopulations (e.g., by age, gender)? NA

Is it possible to identify individuals (i.e., one or more natural persons), either directly or indirectly (i.e., in combination with other data) from the dataset? NA

Does the dataset contain data that might be considered sensitive in any way (e.g., data that reveals racial or ethnic origins, sexual orientations, religious beliefs, political opinions or union memberships, or locations; financial or health data; biometric or genetic data; forms of government identification, such as social security numbers; criminal history)? NA

Were any ethical review processes conducted (e.g., by an institutional review board)? No

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

From the paper:

Is your dataset free of biases? No. There are many kinds of biases that can either be quantified, e.g. geo-location (most images originate from the US and Europe) or camera-model (most images are taken with professional DSLR cameras not easily affordable), there are likely many more biases that this dataset does contain. The only thing that this dataset does not contain are humans and parts of humans, as far as our validation procedure is accurate.

Other Known Limitations

From the paper:

Can you guarantee compliance to GDPR? No, we cannot comment on legal issues.

Additional Information

Dataset Curators

YM. Asano, C. Rupprecht, A. Zisserman and A. Vedaldi.

From the paper:

Who created the dataset (e.g., which team, research group) and on behalf of which entity (e.g., company, institution, organization)? The dataset has been constructed by the research group “Visual Geometry Group” at the University of Oxford at the Engineering Science Department.

Licensing Information

The PASS dataset is available to download for commercial/research purposes under a Creative Commons Attribution 4.0 International License. A complete version of the license can be found here. The whole dataset only contains CC-BY licensed images with full attribution information.

Citation Information

@Article{asano21pass,
author = "Yuki M. Asano and Christian Rupprecht and Andrew Zisserman and Andrea Vedaldi",
title = "PASS: An ImageNet replacement for self-supervised pretraining without humans",
journal = "NeurIPS Track on Datasets and Benchmarks",
year = "2021"
}

Contributions

Thanks to @mariosasko for adding this dataset.

Downloads last month
189

Spaces using pass 2