Relative Uncertainty Learning for Facial Expression Recognition

Authors: Yuhang Zhang, Chengrui Wang, Weihong Deng

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments show that RUL outperforms state-of-the-art FER uncertainty learning methods in both realworld and synthetic noisy FER datasets.
Researcher Affiliation Academia Yuhang Zhang, Chengrui Wang, Weihong Deng Beijing University of Posts and Telecommunications zyhzyh@bupt.edu.cn, crwang@bupt.edu.cn, whdeng@bupt.edu.cn
Pseudocode No No structured pseudocode or algorithm blocks were found in the paper.
Open Source Code Yes The code is available at https://github.com/zyh-uaiaaaa/Relative-Uncertainty-Learning.
Open Datasets Yes RAF-DB [27] is a crowdsourced facial expression dataset that contains 29672 facial images... FER2013 [14] consists of 35,887 grayscale 48x48 pixel images in total... Affect Net [33] is currently the largest FER dataset, including 440,000 images.
Dataset Splits Yes RAF-DB [27] ... 12271 images as training data and 3068 images as test data. FER2013 [14] ... with 28,709 training samples, 3,589 public test samples, and 3,589 private test samples. Affect Net [33] ... around 280,000 training images and 4000 testing images annotated by human.
Hardware Specification Yes The model is trained in an end-to-end manner with a single GTX 1080ti GPU for 70 epochs with batch size of 64.
Software Dependencies No The paper mentions software components like 'Res Net18', 'Adam optimizer', and 'Exponential LR' but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We set dropout rate as 0.4, output dimension as 64. The model is trained in an end-to-end manner with a single GTX 1080ti GPU for 70 epochs with batch size of 64. We also utilize an Adam optimizer [24] with weight decay of 0.0001. The learning rate is initialized as 0.0002 except the last fully connected layer for classification, which is 0.002. We use Exponential LR [30] learning rate scheduler with gamma of 0.9 to decrease the learning rate after each epoch.