[Cross-Age Reference Coding for Age-Invariant Face Recognition and Retrieval

Paper Site: https://bcsiriuschen.github.io/CARC/

Problem Definition

Existing datasets contains little variation in aging. However, faces across age can be very different. Thus it's necessary to solve the challenging problem of face matching with age variation.

Contribution and Discussion

1. Propose a new coding framework called CARC that leverages a reference image set (available from Internet) for age-invariant face recognition and retrieval.

2. Introduce a new large-scale face dataset, CACD, for evaluating face recognition and retrieval across age. The dataset contains more than 160,000 images with 2,000 people and is made publicly available.

3. Conduct extensive experiments on MORPH and CACD and show that CARC can outperform state-of-the-art methods on both datasets.

4. In the future, want to investigate how to effectively choose a subset from the reference people and further improve the performance of age-invariant face recognition and retrieval.

Method

1. Cross-Age Reference Coding (CARC)
   1.1 System Overview

   
   
   System overview of the proposed method

1.2 Reference Set Representation
Using the local features extracted from images of the reference people, the reference set representation is as follows

1.3 Encoding feature into the reference space

1.4 Aggregating representation across different years, proposing to use max pooling to achieve the goal.

2. Cross-Age Celebrity Dataset (CACD)
   2.1 Celebrity Name Collection: Collect names of celebrities whose birth date is from 1951 to 1990. In this 40 years period, we collect the names of top 50 popular celebrities from each birth year with 2,000 names in total.

2.2 Image Collection: Use Google Image Search to collect images with a combination of celebrity name and year as keywords.

2.3 Dataset Statistics: