RankPose: Learning Generalised Feature with Rank Supervision for Head Pose Estimation

Paper

RankPose: Learning Generalised Feature with Rank Supervision for Head Pose Estimation

Abstract

We address the challenging problem of RGB image-based head pose estimation. We first reformulate head pose representation learning to constrain it to a bounded space. Head pose represented as vector projection or vector angles shows helpful to improving. performance. Further, a ranking loss combined with MSE regression loss is proposed. The ranking loss supervises a neural network with paired samples of the same person and penalises incorrect ordering of pose prediction. Analysis on this new loss function suggests it contributes to a better local feature extractor, where features are generalised to Abstract Landmarks which are pose-related features instead of pose-irrelevant information such as identity, age, and lighting. Extensive experiments show that our method significantly outperforms the current state-of-the-art schemes on public datasets:
AFLW2000 and BIWI. Our model achieves significant improvements over previous SOTA MAE on AFLW2000 and BIWI from 4.50 [11] to 3.66 and from 4.0 [24] to 3.71 respectively.

Dependencies

• pytorch >= 0.4.1
• albumentations
• opencv2
• yaml

pip3 install requirements.txt

Datasets

Train data

Face Alignment Across Large Poses: A 3D Solution

300W-LP

Test data

AFLW2000 BIWI Kinect

Train and test

Training

CUDA_VISIBLE_DEVICES=0 python3 train.py ../config/headpose_resnet.yaml

Testing

CUDA_VISIBLE_DEVICES=0 python3 test.py

Pretrained model

Will be available for download in the future.