FreeEnricher: Enriching Face Landmarks without Additional Cost

Authors: Yangyu Huang, Xi Chen, Jongyoo Kim, Hao Yang, Chong Li, Jiaolong Yang, Dong Chen

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate our method, we manually label the dense landmarks on 300W testset. Our method yields state-of-the-art accuracy not only in newly-constructed dense 300W testset but also in the original sparse 300W and WFLW testsets without additional cost.
Researcher Affiliation Industry Yangyu Huang, Xi Chen, Jongyoo Kim*, Hao Yang, Chong Li, Jiaolong Yang, Dong Chen Microsoft Research Asia {yanghuan, xichen6, jongk, haya, chol, jiaoyan, doch}@microsoft.com,
Pseudocode No The paper describes the framework components and their training/testing processes in detail, accompanied by a diagram (Fig. 2), but it does not include any formal pseudocode or algorithm blocks.
Open Source Code No The paper mentions, 'We release the enriched 300W with both preprocessed training set and annotated testing set,' referring to a dataset, but it does not provide any statement about the release of source code for the methodology or a link to a code repository.
Open Datasets Yes To successfully train such deep models, a large number of face alignment datasets have been published, e.g., COFW (Burgos-Artizzu, Perona, and Doll ar 2013), 300W (Sagonas et al. 2013) and WFLW (Wu et al. 2018). Despite of enormous growth in face alignment, there have been very few researches handling dense facial landmarks.
Dataset Splits Yes We newly construct an enriched 300W test dataset by manually labeling the original images, which is employed to evaluate the performance of enriched face alignment. To handle different densities of landmarks, we labeled the continuous curve rather than discrete points. From the curve, firstly, we extract the anchor points by finding the points (on the curve) which are closest to the original landmark labels in 300W. Then, between two neighboring anchor points, the new points are uniformly sampled.
Hardware Specification Yes The Free Enricher model is trained on four GPUs (16GB NVIDIA Tesla P100) by PyTorch (Paszke et al. 2019), where the batch size of each GPU is 68 for 300W and 98 for WFLW.
Software Dependencies Yes The Free Enricher model is trained on four GPUs (16GB NVIDIA Tesla P100) by PyTorch (Paszke et al. 2019), where the batch size of each GPU is 68 for 300W and 98 for WFLW. We employ Adam optimizer with the initial learning rate of 1 10 3 and decay the learning rate by 1/10 at the epochs of 100, 150, and 180, finally ending at 200. Specifically, landmark initializing adopts bspline interpolation method implemented in Scipy (Virtanen et al. 2020) with order of 3 and enriching density of 5, offset generating randomly generates offset by the uniform distribution U( 8, +8), and patch cropping crops 64 64 patch from aligned face image with 1024 1024 resolution (patchface ratio of 1/16) by the center of target landmark.
Experiment Setup Yes The Free Enricher model is trained on four GPUs (16GB NVIDIA Tesla P100) by PyTorch (Paszke et al. 2019), where the batch size of each GPU is 68 for 300W and 98 for WFLW. We employ Adam optimizer with the initial learning rate of 1 10 3 and decay the learning rate by 1/10 at the epochs of 100, 150, and 180, finally ending at 200. Specifically, landmark initializing adopts bspline interpolation method implemented in Scipy (Virtanen et al. 2020) with order of 3 and enriching density of 5, offset generating randomly generates offset by the uniform distribution U( 8, +8), and patch cropping crops 64 64 patch from aligned face image with 1024 1024 resolution (patchface ratio of 1/16) by the center of target landmark.