Article
Computer Science, Artificial Intelligence
Zhenfeng Fan, Xiyuan Hu, Chen Chen, Xiaolian Wang, Silong Peng
Summary: Global dense registration of 3D faces commonly relies on facial landmark correspondences. However, accurately annotating landmarks in raw 3D face scans is not always easy. This study proposes a general framework without pre-annotated landmarks, which improves robustness and allows uniform mesh deformation. The framework includes two stages and revisits dense registration from semantic and topological perspectives. Experimental evaluation demonstrates its effectiveness in handling noisy, occluded, and highly deformed data.
PATTERN RECOGNITION
(2023)
Article
Computer Science, Artificial Intelligence
Zhenfeng Fan, Silong Peng, Shihong Xia
Summary: This paper proposes an iterative dividing and diffusing method to achieve an optimum solution for dense vertex-to-vertex correspondence between 3D faces. The method involves a local registration problem for dividing and a linear least-square problem for diffusing, with constraints on fixed features and a multi-resolution algorithm for acceleration. The proposed method is validated through extensive experiments on public datasets, showing improved representations of 3D facial data and coherent local deformations with elegant grid architecture.
INTERNATIONAL JOURNAL OF COMPUTER VISION
(2023)
Article
Computer Science, Artificial Intelligence
Romain Belmonte, Benjamin Allaert, Pierre Tirilly, Ioan Marius Bilasco, Chaabane Djeraba, Nicu Sebe
Summary: In this paper, the impact of facial landmark localization (FLL) approaches on facial expression recognition (FER) tasks is studied. Due to the limitations of existing FLL datasets in measuring performance under different difficulties, the performance of recent approaches is quantified under variations in head pose and facial expressions. The study shows that optimizing the euclidean distance for landmark accuracy does not necessarily improve FER performance. To address this issue, a new evaluation metric for FLL that is more relevant to FER is proposed.
IEEE TRANSACTIONS ON AFFECTIVE COMPUTING
(2023)
Article
Chemistry, Analytical
Junsik Kim, Hyungwha Jeong, Jeongmok Cho, Changsik Pak, Tae Suk Oh, Joon Pio Hong, Soonchul Kwon, Jisang Yoo
Summary: Treatment of facial palsy is essential to avoid serious sequelae and further damage. This study introduces three numerical methods to measure the degree of facial palsy and compare it with previous images. Experiments confirm that the proposed numerical approach is beneficial for assessing the progression of facial palsy.
Article
Engineering, Electrical & Electronic
Zhiqun Pan, Yongxiong Wang, Sunjie Zhang
Summary: This paper proposes a real-time framework for joint face detection and Facial Landmark Localization (FLL). It utilizes a fully convolutional network to predict the location of facial landmarks and face regions, and introduces a progressively pseudo labeling training method to eliminate the effect of inaccurate/noisy annotations. Two graph matching algorithms without learnable parameters are also proposed for completing the bottom-up face assembly.
SIGNAL PROCESSING-IMAGE COMMUNICATION
(2022)
Article
Engineering, Electrical & Electronic
Xin Jin, Zhonglan Li, Ning Ning, Huimin Lu, Xiaodong Li, Xiaokun Zhang, Xingfan Zhu, Xi Fang
Summary: The image-based virtual illumination technology enables illumination swapping between two face images and re-adjustment of the overall illumination while preserving geometric features.
IEEE SENSORS JOURNAL
(2022)
Article
Computer Science, Artificial Intelligence
Congcong Zhu, Xiaoqiang Li, Jide Li, Songmin Dai, Weiqin Tong
Summary: This paper proposes a structural relation network (SRN) for occlusion-robust landmark localization, which aims to capture the structural relations among different facial components to improve localization accuracy. The method considers spatial relations in model design, enhances training data by synthesizing occluded faces, and extends the ability to handle occluded video data using a Markov decision process.
PATTERN RECOGNITION
(2022)
Article
Veterinary Sciences
Van Nguyen, Luis F. Alves F. Pereira, Zhihua Liang, Falk Mielke, Jeroen Van Houtte, Jan Sijbers, Jan De Beenhouwer
Summary: This paper presents an image processing method for studying the 3D musculoskeletal motion of animals. The method consists of two modules: an automated 3D landmark extraction technique and a deep neural network for 2D landmark detection. Experimental results demonstrate that the proposed method can accurately extract landmarks from X-ray images and infer the 3D poses of animals.
FRONTIERS IN VETERINARY SCIENCE
(2022)
Article
Computer Science, Artificial Intelligence
Claudio Ferrari, Stefano Berretti, Pietro Pala, Alberto Del Bimbo
Summary: The 3D Morphable Model (3DMM) is a statistical tool for representing 3D face shapes. This manuscript presents an automatic approach that utilizes 3DMM to establish dense correspondence across 3D faces, accurately fit unseen faces, and transfer semantic annotations. The approach shows promising results in terms of generalization and accuracy in establishing dense correspondence, even with complex facial expressions.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
(2022)
Article
Computer Science, Information Systems
Shinfeng D. Lin, Paulo E. Linares Otoya
Summary: This paper presents a novel approach to achieve pose-invariant face recognition using ensemble learning and local feature descriptors. The proposed method extracts feature vectors from image regions surrounding specific facial landmarks, and trains three different classification models as base learners. Experimental results show better performance compared to existing methods using the CMU-PIE dataset.
Article
Computer Science, Software Engineering
Jie Xiu, Xiujie Qu, Haowei Yu
Summary: In this study, a double discriminative face super-resolution network (DDFSRNet) is proposed to balance the visual quality and pixel accuracy of face super-resolution. The collaborative generator and two discriminators are used to reconstruct facial key components and judge the authenticity of the generated data. The proposed method outperforms other advanced methods in perceptual effect and fitting high-resolution facial images.
Article
Computer Science, Information Systems
Seongmin Lee, Hyunse Yoon, Sohyun Park, Sanghoon Lee, Jiwoo Kang
Summary: This study introduces neural networks to reconstruct stable and precise 3D faces by learning facial changes caused by identity, expression, and temporal cues. The proposed facial alignment network exhibits reliable and precise performance in reconstructing static and dynamic faces.
Article
Computer Science, Artificial Intelligence
Tiancheng Wen, Zhonggan Ding, Yongqiang Yao, Yaxiong Wang, Xueming Qian
Summary: In this paper, a lightweight cascaded facial landmark detector called PicassoNet is proposed to achieve a balance between accuracy and inference speed. PicassoNet integrates refinement submodules into a single network using group convolution to adaptively compute individual facial parts. Additionally, a boundary-aware loss is introduced to improve the predicted positions of keypoints. Experimental results demonstrate that the proposed method outperforms existing methods in terms of both speed and accuracy.
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS
(2023)
Article
Computer Science, Artificial Intelligence
Boyu Chen, Wenlong Guan, Peixia Li, Naoki Ikeda, Kosuke Hirasawa, Huchuan Lu
Summary: The paper introduces a residual multi-task learning framework for predicting facial landmark localization and expression recognition simultaneously, utilizing a residual learning module to strengthen the connection between the two tasks and addressing the lack of training data with multi-task labels in multi-task learning.
PATTERN RECOGNITION
(2021)
Article
Mathematics
Xuxin Lin, Yanyan Liang
Summary: This paper proposes a region-aware deep feature-fused network (RDFN) to address the issue of input initialization in facial landmark localization. The network consists of a region detection subnetwork and a region-wise landmark localization subnetwork, utilizing cross-task feature fusion and within-task feature fusion to extract multi-semantic region features. A location reweighting strategy is employed at the inference stage to obtain 2D landmark coordinates. Experimental results demonstrate the competitive performance of this method on multiple datasets.