Geometric Order Learning for Rank Estimation

Authors: Seon-Ho Lee, Nyeong Ho Shin, Chang-Su Kim

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on facial age estimation, historical color image (HCI) classification, and aesthetic score regression demonstrate that GOL constructs effective embedding spaces and thus yields excellent rank estimation performances. The source codes are available at https://github.com/seon92/GOL
Researcher Affiliation Academia Seon-Ho Lee Nyeong-Ho Shin Chang-Su Kim School of Electrical Engineering Korea University seonholee@mcl.korea.ac.kr, nhshin@mcl.korea.ac.kr, changsukim@korea.ac.kr
Pseudocode No The paper describes the proposed algorithm (GOL) in detail, but it does not include a formal pseudocode block or a clearly labeled algorithm figure.
Open Source Code Yes The source codes are available at https://github.com/seon92/GOL
Open Datasets Yes We use four datasets of MORPH II (Ricanek & Tesafaye, 2006), CACD (Chen et al., 2015), UTK (Zhang et al., 2017), and Adience (Levi & Hassner, 2015), as detailed in Appendix D.3.
Dataset Splits No Table 3 compares the results on CACD, which is a bigger dataset containing over 100,000 natural face shots in diverse environments. GOL outperforms the second-best methods with meaningful gaps of 0.12 and 0.17 in the train and validation settings, respectively, which indicates that it can cope with large and diverse data effectively as well.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments.
Software Dependencies No We initialize an encoder h with VGG16 pre-trained on ILSVRC2012 (Deng et al., 2009) and reference points with the Glorot normal method (Glorot & Bengio, 2010). The Adam optimizer (Kingma & Ba, 2015) is used with a batch size of 32 and a weight decay of 5 × 10−4, and the initial learning rates for the encoder and the reference points are set to 10−4 and 10−3, respectively. We perform the scheduled learning according to cosine annealing cycles (Huang et al., 2017).
Experiment Setup Yes We initialize an encoder h with VGG16 pre-trained on ILSVRC2012 (Deng et al., 2009) and reference points with the Glorot normal method (Glorot & Bengio, 2010). The Adam optimizer (Kingma & Ba, 2015) is used with a batch size of 32 and a weight decay of 5 × 10−4, and the initial learning rates for the encoder and the reference points are set to 10−4 and 10−3, respectively. We perform the scheduled learning according to cosine annealing cycles (Huang et al., 2017).