AGRNet: Adaptive Graph Representation Learning and Reasoning for Face Parsing

Te, Gusi; Hu, Wei; Liu, Yinglu; Shi, Hailin; Mei, Tao

doi:10.1109/tip.2021.3113780

Cited by 24 publications

(8 citation statements)

References 49 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…necklace in CelebAMask-HQ). Even though, the required resolution of FaRL 448 ft is lower than the resolution needed by the state-of-the-art approach [94] which is 473. It is also worth noting that the backbone-frozen FaRL achieves even better performances than the Scratch models on both benchmarks, showing that the representation learned from FaRL is not only widely applicable, but also sufficiently effective.…”

Section: Comparing With State-of-the-art Face Methodsmentioning

confidence: 93%

“…The refined FaRL ft brings even higher F1 scores. The ultimate performance is achieved by FaRL 448 ft , which outperforms the state-of-the-art method [94] by 1.58 and 4.06 on LaPa and CelebAMask-HQ, respectively. We note that the input resolution plays a critical role in face parsing performance, it is especially effective for small components (e.g.…”

Section: Comparing With State-of-the-art Face Methodsmentioning

confidence: 94%

“…In the area of face analysis, most tasks [7,42,50,57,94,95] are solved by supervised training with manually labeled data. Such supervised methods require a large number of training samples and may suffer from overfitting because of the massive number of model parameters.…”

Section: Facial Representation Learningmentioning

confidence: 99%

“…CelebAMask-HQ consists of around 30K facial images, 24,183 for training and 2,824 for testing, each annotated with a 19-category label map including not only facial components but also body parts and accessories like eyeglasses, earrings and necklaces. Following [58,94,95], the F1 scores of facial components are used to measure the performance. Face alignment aims to regress 2D face landmark coordinates on a face image.…”

Section: Downstream Face Tasksmentioning

confidence: 99%

“…Face analysis tasks are of crucial importance to social interactions and have received extensive attention over the past decades. Many existing state-of-the-art results [7,50,94] come from deep neural networks with supervised learning. However, such supervised models, in order to learn appropriate feature representations for each given task, are studied separately with large-scale manually annotated data which is expensive and difficult to acquire, especially for some face tasks such as face parsing and face alignment.…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

General Facial Representation Learning in a Visual-Linguistic Manner

Zheng¹,

Yang²,

Zhang³

et al. 2021

Preprint

View full text Add to dashboard Cite

How to learn a universal facial representation that boosts all face analysis tasks? This paper takes one step toward this goal. In this paper, we study the transfer performance of pre-trained models on face analysis tasks and introduce a framework, called FaRL, for general Facial Representation Learning in a visual-linguistic manner. On one hand, the framework involves a contrastive loss to learn high-level semantic meaning from image-text pairs. On the other hand, we propose exploring low-level information simultaneously to further enhance the face representation, by adding a masked image modeling. We perform pre-training on LAION-FACE, a dataset containing large amount of face image-text pairs, and evaluate the representation capability on multiple downstream tasks. We show that FaRL achieves better transfer performance compared with previous pretrained models. We also verify its superiority in the lowdata regime. More importantly, our model surpasses the state-of-the-art methods on face analysis tasks including face parsing and face alignment.* Equal contribution. † Work done during internship at MSRA.

show abstract

Section: Comparing With State-of-the-art Face Methodsmentioning

confidence: 93%

Section: Comparing With State-of-the-art Face Methodsmentioning

confidence: 94%