The 4th Recognizing Families In the Wild (RFIW)

Source: Deep Learning on Medium

What is new for RFIW-2020?

The new components of RFIW2020 are listed as follows:

  • Two additional tracks, along with the return of kinship verification (i.e., three tracks in total).
  • General call for papers in work in automatic kinship recognition.
  • A Brave New Idea track is calling for innovative ways of viewing the problem.

New challenges!

Along with the traditional verification task, RFIW2020 will support two new tracks, tri-subject verification, and large-scale search-and-retrieval.

Tri-subject verification (T-2) focuses on a slightly different view of kinship verification — the goal is to decide whether a child is related to a pair of parents. This paradigm follows a 2-to-1 verification protocol with all rules mimicking that of verification, with the difference being one item from each pair consisting of a man and woman, and the question then is “are these the parents.” Tri-subject is a natural extension of the verification task, as it is a more realistic assumption, as knowing one parent typically means information of the other is accessible.

Large-scale search-and-retrieval (T-3) will mimic that of template-based, open-sets protocols provided by benchmarks like IJB-B. A gallery of distractors and real relatives, with the task of ranking, faces are then sorted by the likelihood of being kin (i.e., blood relative). T-3 closely mimics the real-world application of missing children. For instance, a child is found online, exploited by the unknown, and it is unlikely not in any database; however, a family member likely is– identify a family member, determine the identity of the unknown. Additionally, reuniting families split from the modern-day refugee crisis. Provided technology to recognize family members via visual media, we could then match families together from different camps at the cost of a low-cost security video-feed.

Benchmarks, along with results of prior RFIW, will be provided; also, source code to reproduce and demonstrate each task end-to-end will be made available. Thus, enabling newcomers while challenging the experts. We will also call for general paper submissions of new work to expand the types of problems and the use-case of the FIW dataset.

Call for papers!

In addition to the three organized task evaluations, we will also add this piece to RFIW2020 (i.e., papers that use FIW in novel ways). The main reason we added this is to challenge researchers to propose novel technology besides the task evaluations. We found the assessments to be great for structuring existing problems such that researchers and practitioners can make fair comparisons of algorithms. However, this limits the scope of the problems of automatic kinship recognition. From this, we expect the light to shed on one or more of the following ways:

  • To advance the state-of-the-art for kinship verification and family classification.
  • To benchmark new tasks for FIW, like fine-grain classification, large-scale search, and retrieval, tri-subject verification.
  • To propose generative models for family photos, relative faces, photo albums, such.
  • To explore and understand multimodal uses of text captions accompanying the family photos of FIW.
  • To pitch cluster, multi-view, and various types of problems.
  • To treat kinship as a soft attribute for higher-level tasks (e.g., facial recognition, group understanding, social media analysis).
  • Much more.