Sample and Computation Redistribution for Efficient Face Detection

Authors: Jia Guo, Jiankang Deng, Alexandros Lattas, Stefanos Zafeiriou

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments conducted on WIDER FACE demonstrate the state-of-the-art accuracy-efficiency trade-off for the proposed SCRFD family across a wide range of compute regimes.
Researcher Affiliation Collaboration Jia Guo2, Jiankang Deng1,2 , Alexandros Lattas1,3, Stefanos Zafeiriou1,3 1Huawei, 2Insight Face, 3Imperial College London {guojia,jiankangdeng}@gmail.com, {a.lattas,s.zafeiriou}@imperial.ac.uk
Pseudocode Yes Algorithm 1: Search algorithm for computation redistribution
Open Source Code Yes Code is available at: https://github.com/deepinsight/insightface/ tree/master/detection/scrfd.
Open Datasets Yes train each model on the WIDER FACE (Yang et al., 2016) training set for 80 epochs.
Dataset Splits Yes The WIDER FACE dataset is split into training (40%), validation (10%) and testing (50%) subsets by randomly sampling from 61 scene categories.
Hardware Specification Yes Our experiments are implemented in Py Torch... and train on eight Tesla V100.
Software Dependencies No Our experiments are implemented in Py Torch, based on the open-source MMDetection (Chen et al., 2019a). However, specific version numbers for PyTorch or MMDetection are not provided.
Experiment Setup Yes We adopt the SGD optimizer (momentum 0.9, weight decay 5e-4) with a batch size of 8 × 8 and train on eight Tesla V100. The learning rate is linearly warmed up to 0.015 within the first 3 epochs. During network search, the learning rate is multiplied by 0.1 at the 55-th, and 68-th epochs. The learning process terminates on the 80-th epoch. For training of both baselines and searched configurations, the learning rate decays by a factor of 10 at the 440-th and 544-th epochs, and the learning process terminates at the 640-th epoch.