Order Learning and Its Application to Age Estimation

Authors: Kyungsun Lim, Nyeong-Ho Shin, Young-Yoon Lee, Chang-Su Kim

ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We apply order learning to develop a facial age estimator, which provides the state-of-the-art performance. Moreover, the performance is further improved when the order graph is divided into disjoint chains using gender and ethnic group information or even in an unsupervised manner.
Researcher Affiliation Collaboration Kyungsun Lim, Nyeong-Ho Shin, Young-Yoon Lee, and Chang-Su Kim School of Electrical Engineering, Korea University and Samsung Electronics Co., Ltd {kslim, nhshin, cskim}@mcl.korea.ac.kr, yy77lee@gmail.com
Pseudocode Yes Algorithm 1 Order Learning with Unsupervised Chains Algorithm 2 Regular Assign(K, X, L)
Open Source Code No The paper does not provide any links or explicit statements about the availability of open-source code for the described methodology.
Open Datasets Yes MORPH II (Ricanek & Tesafaye, 2006) is the most popular age estimation benchmark, containing about 55,000 facial images in the age range [16, 77]. IMDB-WIKI (Rothe et al., 2018) is another dataset containing about 500,000 celebrity images obtained from IMDB and Wikipedia. ... we form a balanced dataset from MORPH II, AFAD (Niu et al., 2016), and UTK (Zhang et al., 2017b).
Dataset Splits Yes The balanced dataset is partitioned into training and test subsets with ratio 8 : 2. For performance assessment, we calculate the mean absolute error (MAE) (Lanitis et al., 2004) and the cumulative score (CS) (Geng et al., 2006). ... As evaluation protocols for MORPH II, we use four different settings, including the 5-fold subject-exclusive (SE) and the 5-fold random split (RS) (Chang et al., 2010; Guo & Wang, 2012).
Hardware Specification No The paper does not specify any particular hardware (e.g., GPU model, CPU, cloud instance type) used for running experiments.
Software Dependencies No The paper mentions software like VGG16, Image Net, Adam optimizer, Seeta Face Engine, but does not provide specific version numbers for these or any other software dependencies, which is required for reproducibility.
Experiment Setup Yes We align all facial images using Seeta Face Engine (Zhang et al., 2014) and resize them into 256 × 256 × 3. Then, we crop a resized image into 224 × 224 × 3. For the feature extractors in Figure 2, we use VGG16 without the FC layers (Simonyan & Zisserman, 2014). They yield 512-channel feature vectors. Then, the vectors are concatenated and input to the ternary classifier, which has three FC layers, yielding 512-, 512-, and 3-channel vectors sequentially. ... In (10)–(12), τage is set to 0.1. ... We set the learning rate to 10−4 for the first 70 epochs. Then, we select 5 references for each age class. Using the selected references, we fine-tune the network with a learning rate of 10−5. We repeat the reference selection and the parameter fine-tuning up to 3 times.