Fashion 10000: An Enriched Dataset of Fashion and Clothing
?Download as PPTX, PDF?
6 likes?2,234 views
The document summarizes a dataset of over 10,000 fashion and clothing images collected from Flickr and annotated with metadata and labels. Key details include: - The dataset contains images from 262 fashion categories with on average 122 photos per category. - Images were annotated on Amazon Mechanical Turk based on 6 labels regarding fashion relevance, number of people, professional models, and formality. - Over 24,000 assignments were completed with high inter-annotator agreement for most labels. - The dataset aims to support applications in social media analysis, benchmarking tasks, and studying intentional image framing.
Fashion 10000: An Enriched Dataset of Fashion and Clothing
- 1. Fashion 10000 An Enriched Dataset of Fashion and Clothing Presentation: Michael Riegler, Klagenfurt University & TU Delft Babak Loni, TU Delft Lei Yen Cheung, TU Delft Alessandro Bozzon, TU Delft Luke Gottlieb, ICSI Martha Larson, TU Delft
- 2. Table of Content ? Introduction ? Dataset Collection ? Dataset Annotation C Statistics ? Applications of Dataset ? Conclusion
- 3. The Dataset ? Social Images ? At least 10000 fashion- related images ? Social metadata ? Creative Common images ? Annotated with different labels
- 4. The Collection Wikipedia 470 Fashion Categories Flickr - Query only CC attribution images - Query should also appear in tags - Top relevant images 32K Images 262 Categories Flickr Fashion 10000 + MTurk Annotations + Metadata
- 5. Metadata ? Collected in xml and csv format C Title, description, owner, Tags, Location, geo-parameters ? Additional metadata: Info, Geos, Context, Tags, Notes, Favorites, Urls, Comments
- 6. General Statistics Pairs fashion item, photo 32,398 Number of distinct fashion categories 262 Max/avg/min nr of photos per fashion item 200/ 122.95 / 10 Number of photos with geo annotations 7,933 Total number of comments 58,578 Max/avg/min nr of comments per photo 575 / 7.35/ 1 Total number of tags, photo pairs 460,907 Total number of distinct tags 56,275 Max/avg/min nr of tags per photo 136/ 15.15/ 1 Total number of notes, photo pairs 5,892 Max/avg/min nr of notes per photo 195/ 5.31/ 1 Total number of favorites 37,131 Max/avg/min nr of favorites per photo 20/ 3.61/ 1 Total number of contexts 110,505 Max/avg/min nr of contexts per photo 206/ 3.93/ 1
- 7. General Statistics Pairs fashion item, photo 32,398 Number of distinct fashion categories 262 Max/avg/min nr of photos per fashion item 200/ 122.95 / 10 Number of photos with geo annotations 7,933 Total number of comments 58,578 Max/avg/min nr of comments per photo 575 / 7.35/ 1 Total number of tags, photo pairs 460,907 Total number of distinct tags 56,275 Max/avg/min nr of tags per photo 136/ 15.15/ 1 Total number of notes, photo pairs 5,892 Max/avg/min nr of notes per photo 195/ 5.31/ 1 Total number of favorites 37,131 Max/avg/min nr of favorites per photo 20/ 3.61/ 1 Total number of contexts 110,505 Max/avg/min nr of contexts per photo 206/ 3.93/ 1
- 8. General Statistics Pairs fashion item, photo 32,398 Number of distinct fashion categories 262 Max/avg/min nr of photos per fashion item 200/ 122.95 / 10 Number of photos with geo annotations 7,933 Total number of comments 58,578 Max/avg/min nr of comments per photo 575 / 7.35/ 1 Total number of tags, photo pairs 460,907 Total number of distinct tags 56,275 Max/avg/min nr of tags per photo 136/ 15.15/ 1 Total number of notes, photo pairs 5,892 Max/avg/min nr of notes per photo 195/ 5.31/ 1 Total number of favorites 37,131 Max/avg/min nr of favorites per photo 20/ 3.61/ 1 Total number of contexts 110,505 Max/avg/min nr of contexts per photo 206/ 3.93/ 1
- 9. General Statistics Pairs fashion item, photo 32,398 Number of distinct fashion categories 262 Max/avg/min nr of photos per fashion item 200/ 122.95 / 10 Number of photos with geo annotations7,933 Total number of comments 58,578 Max/avg/min nr of comments per photo 575 / 7.35/ 1 Total number of tags, photo pairs 460,907 Total number of distinct tags 56,275 Max/avg/min nr of tags per photo 136/ 15.15/ 1 Total number of notes, photo pairs 5,892 Max/avg/min nr of notes per photo 195/ 5.31/ 1 Total number of favorites 37,131 Max/avg/min nr of favorites per photo 20/ 3.61/ 1 Total number of contexts 110,505 Max/avg/min nr of contexts per photo 206/ 3.93/ 1
- 10. Dataset Annotation ? Some images might not be relevant to fashion and clothing ? The ground truth differentiates relevant from non-relevant
- 11. Dataset Annotation ? We used AMT to create ground truth for the images ? The fashion category is described with a definition from Wikipedia ? 6 questions to create 6 labels for each of the images ? We also ask about familiarity of workers with the fashion category
- 12. HIT Design
- 13. HIT Design
- 14. HIT Questions (Labels) Question Possible values Q1) Fashion / Clothing Related yes C no - notsure Q2) Specialty clothing item (image Category) yes C no - notsure Q3) Number of people nopeople C onepeople - manypeople Q4) Professional model or not? yes C no C notapp (not applicable) Q5) Person wearing fashion? yes C no C noperson C notapp (not applicable) Q6) Formal / Informal formalmen - formalwomen - informalmen informalwomen C other (cross-dressing or multiple persons) C notapp (not applicatble)
- 15. Annotation Statistics Total number of assignments 24,457 % of rejected assignments 4 % Total number of unique workers 1470 Avg. number of assignment by each worker 17 Avg. Completion time 127 sec Avg. familiarity of workers with fashion items 5.8 (range 1-7) Question 1 2 3 4 5 6 Kappa Value 0.66 0.65 0.85 0.51 0.38 0.48
- 16. Dataset Statistics ? Using the generated ground truth the statistics about the images were calculated Number of fashion related images 18,487 Number of images with many people 7,417 Number of images with one person 9,771 Number of images with no person 13,179 Number of images with intention of showing fashion 9,096 Number of professional fashion images 2,814
- 17. Applications of the Dataset ? Developing social media content analysis C Game with a purpose (domino game) ? Basis for the brave new task in MediaEval multimedia benchmarking initiative ? Use case for the proof of intentional framing
- 18. Conclusion ? Fashion dataset ? Six different labels ? AMT generated ground truth ? Can be used in various research areas ? Evaluated in the MediaEval Benchmark
Editor's Notes
- Social Images: From user-generated fashion categories to user-generated images Should include at least 10000 fashion-related images Should include social metadata Only Creative Common Attribution images Images should be annotated with different labels
- Collected in xml and csv format General information about images are available in xml format Title, description, owner,.. Tags Location, geo-parameters Additional metadata were collected using different Flickr APIs and converted to csv format Info, Geos, Comments, Context, Tags, Notes, Favorites, Urls
- Some images might not be relevant to fashion and clothing The ground truth for the images differentiates relevant from non-relevant images Depending on the usage of dataset different type of annotations might be necessary
- We used AMT to create ground truth for the images In each HIT we ask workers to annotate 4 images from one category The fashion category is described to workers with a definition from Wikipedia In total we ask 6 questions to create 6 labels for each of the images We also ask about familiarity of workers with the Fashion Category The range of value can be from 1 (unfamiliar) to 7 (familiar) Different options are explained with visual popups In addition to the csv file generated by AMT, we created another csv to represent ground truth Each of the images are annotated by 3 workers The ground truth are generated by majority voting
- Kappa statistics have been used to calculate the agreement among annotators For each of the six questions in the HIT, the agreement among three workers was calculated separately General statistics about the workers and the hit