Paper Track


Topics Covered

Topics of the papers include but are not limited to:

  • ML Art Galleries and ML Music Festivals retrospectives

  • Understanding bias and fairness in the context of image generation. Many contemporary artists have begun to engage with AI and ML. Some have done this for the visual aesthetics that the techniques begin to allow, others engage with them more critically in order to understand and reveal the algorithmic processes that are beginning to have great social and political power. These groups often notice the social dimension of the algorithms which can be overlooked by the computer science community. What biases does the choice of training set bring to the images that are created? In a domain of subjective evaluation, what biases does the researcher themselves bring to determining what algorithms are a success?

  • Better approaches for real-time interactive methods for design. Current image generation algorithms show its value in efficiently generating content with various styles. This capability combined with designers' creativity and understanding of human emotions can potentially induce a new form of design process, where designers interact with the algorithms in real-time, leading to more valuable design outcomes.

  • Artistic perspectives on existing technologies and tools. The interdisciplinary arts brings an alternative perspective to computer vision work that can help us see the field in terms of its broader social context. Artists are often among the first groups to explore unintended consequences of new technologies, and to challenge foundational assumptions of the fields with which they interact.

  • Risks of cultural appropriation. While ethical considerations for what concerns the work of a specific artist are starting to be discussed in depth, the aspect of broader cultural appropriation is still relatively unexplored. Fundamental challenges arise when trying to define ownership and copyrights in the context of Traditional Cultural Expression, where the intellectual contribution can't be attributed to a single individual, but results from, and often defines, the cultural evolution of specific groups of people.

  • Copyright, ownership, and licensing. Generating creative content through generative models involves using already existing content as input which serves as an example of the desired output. As a result, this raises questions about style origination and ownership. Generated content tends to carry some resemblance the content used as examples in the generation process. Could these similarities be interpreted as an imitation of style or merely drawing inspiration from existing work and where do we draw the line?

  • Considerations relating to dataset diversity. Fashion and Art are two ways in which we show ourselves to the world. They can also be understood as a form of expression and storytelling. They have proven over the years to be efficient in preservation and celebration of culture and history. Existing work on Fashion and Art synthesis in computer vision research have been only limited to the western forms. This comes as a result of lack of diversity in the available datasets which further perpetuates the danger of a single story and under representation of non-western cultures and Art forms. We invite submissions that analyze the bias aspect of existing datasets as well as new diverse benchmarks for fashion and art (e.g. [5, 6]).

  • Applications for users with diverse preferences. Computer vision applications in fashion or other creative domains which consider the diversity of user preferences, needs and requirements, such as individual tastes, body types [3, 4], aesthetic needs, etc.

  • Eco-friendly applications that help reduce waste. Computer vision applications in fashion or other creative domains which aim at reducing wastes, and eco-friendly recommendations, such as suggesting minimal quantities of fashion items to build a minimalist wardrobe [1].

  • Influence in fashion and artistic domain. Approaches that model influence between people, its impact on the fashion and creative domains, and how influence strength relates to other factors like cultural background, social and financial status [7].

  • Contributions that address the short time span of the research data and its implications on developed models. Most datasets for creative arts and fashion have limited temporal coverage that tend to range from a few months to a couple of years. However, certain concepts like cultural appropriation and artwork attribution may span a time range of several years or decades which may create serious challenges in pinpointing the original creators.

  • Robustness testing in fashion retrieval systems.

References:

[1] W.-L. Hsiao and K. Grauman, “Creating capsule wardrobes from fashion images,” in CVPR, 2018

[2] B. Zhao, J. Feng, X. Wu, and S. Yan, “Memory-augmented attribute manipulation networks for interactive fashion search,” in CVPR, 2017.

[3] S. C. Hidayati, T. W. Goh, J. G. Chan, C. Hsu, J. See, W. Lai Kuan, K. Hua, Y. Tsao, and W. Cheng, “Dress with style: Learning style from joint deep embedding of clothing styles and body shapes,” IEEE Transactions on Multimedia, 2020.

[4] W-L. Hsiao and K. Grauman. "ViBE: Dressing for Diverse Body Shapes," in CVPR 2020

[5] H. Kataoka, Y. Satoh, K. Abe, M. Minoguchi, and A. Nakamura, “Ten-million-order human database for world-wide fashion culture analysis,” in CVPR Workshops, 2019.

[6] K. Matzen, K. Bala, and N. Snavely, “Streetstyle: Exploring world-wide clothing styles from millions of photos,” in arXiv, 2017.

[7] Z. Al-Halah and K. Grauman, “From Paris to Berlin: Discovering Fashion Style Influences Around the World,” in CVPR, 2020.

Submission and Presentation Guidelines

We solicit short papers on developing and applying computer vision techniques that are valuable in the creative domains, with an emphasis on fashion, art and design. Accepted short papers will be linked online at the workshop webpage. The page limit is between two to eight pages (excluding references). We encourage submission of work that has been previously published, and work in progress on relevant topics of the workshop. Accepted papers will be presented at the poster session. Manuscripts should follow the CVPR 2022 template and should be submitted through our CMT portal.

Paper submission Link: https://cmt3.research.microsoft.com/EC3V2022

Note to Authors

  • For authors who want to submit their accepted work at this workshop to a different journal or conference, please check their double submission rules thoroughly. The four page limit should comply with the most common conferences on computer vision. However, the authors need to confirm this with the venue they intend to submit the paper to.

  • Please note that our workshop doesn't have a proceeding.

  • The review process is single-blind.

  • Authors can optionally submit supplemental materials for the paper via CMT.

  • For poster session, the poster dimension requirement is the same as the main conference.

Important Dates

  • Paper submission deadline: April 16 April 23 (11:59PM EST)

  • Notification to authors: April 30 May 7 (11:59PM EST)