Graduate Thesis Or Dissertation
 

Adaptive Wing Loss for Robust Face Alignment via Heatmap Regression

Public Deposited

Downloadable Content

Download PDF
https://ir.library.oregonstate.edu/concern/graduate_thesis_or_dissertations/nk322k521

Descriptions

Attribute NameValues
Creator
Abstract
  • Heatmap regression has became one of the mainstream approaches to localize facial landmarks. As Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) are becoming popular in solving computer vision tasks, extensive research has been done on these architectures. However, the loss function for heatmap regression is rarely studied. In this paper we present two issues with common approaches for heatmap regression: i) The widely used MSE loss is not able to reduce small regression errors. ii) Equal weights are assigned on a very unbalanced background and foreground pixels on heatmap. To resolve the first issue, we analyze the ideal loss function properties that are desired for heatmap regression in face alignment problems. Then we propose a novel loss function, named Adaptive Wing loss, that is able to adapt its shape to ground truth heatmap pixel intensities. This adaptability allows small to medium range errors to have a strong influence on foreground pixels and weak influence on background pixels for a more accurate regression result. The second issue is resolved with our proposed Weighted Loss Map, which assigns high weights on foreground and difficult background pixels to help training process focus more on pixels that are crucial to landmark localization. To further improve face alignment accuracy, we introduce boundary prediction and CoordConv with boundary coordinates into our model. Extensive experiments on different benchmarks, including AFLW, 300W and WFLW, show our approach outperforms the state-of-the-arts by a significant margin on various evaluation metrics.
License
Resource Type
Date Issued
Degree Level
Degree Name
Degree Field
Degree Grantor
Commencement Year
Advisor
Committee Member
Academic Affiliation
Subject
Rights Statement
Publisher
Peer Reviewed
Language
Embargo reason
  • Ongoing Research
Embargo date range
  • 2019-01-07 to 2019-06-23

Relationships

Parents:

This work has no parents.

In Collection:

Items

Thumbnail Title Date Uploaded Visibility Actions
file details WangXinyao2019.pdf 2019-01-01 Public Download