Skip to content
Switch branches/tags


[CVPR19] FSA-Net: Learning Fine-Grained Structure Aggregation for Head Pose Estimation from a Single Image

Code Author: Tsun-Yi Yang


  • 2019/10/06: Big thanks to Kapil Sachdeva again!!! The keras Lambda layers are replaced, and converted tf frozen models are supported!
  • 2019/09/27: Refactoring the model code. Very beautiful and concise codes contributed by Kapil Sachdeva.
  • 2019/08/30: Demo update! Robust and fast SSD face detector added!

Comparison video

(Baseline Hopenet:

(New!!!) Fast and robust demo with SSD face detector (2019/08/30)

Webcam demo

Signle person (LBP) Multiple people (MTCNN)
Time sequence Fine-grained structure




Paper authors

Tsun-Yi Yang, Yi-Ting Chen, Yen-Yu Lin, and Yung-Yu Chuang


This paper proposes a method for head pose estimation from a single image. Previous methods often predicts head poses through landmark or depth estimation and would require more computation than necessary. Our method is based on regression and feature aggregation. For having a compact model, we employ the soft stagewise regression scheme. Existing feature aggregation methods treat inputs as a bag of features and thus ignore their spatial relationship in a feature map. We propose to learn a fine-grained structure mapping for spatially grouping features before aggregation. The fine-grained structure provides part-based information and pooled values. By ultilizing learnable and non-learnable importance over the spatial location, different variant models as a complementary ensemble can be generated. Experiments show that out method outperforms the state-of-the-art methods including both the landmark-free ones and the ones based on landmark or depth estimation. Based on a single RGB frame as input, our method even outperforms methods utilizing multi-modality information (RGB-D, RGB-Time) on estimating the yaw angle. Furthermore, the memory overhead of the proposed model is 100× smaller than that of previous methods.


  • Keras
  • Tensorflow
  • GTX-1080Ti
  • Ubuntu
python                    3.5.6                hc3d631a_0  
keras-applications        1.0.4                    py35_1    anaconda
keras-base                2.1.0                    py35_0    anaconda
keras-gpu                 2.1.0                         0    anaconda
keras-preprocessing       1.0.2                    py35_1    anaconda
tensorflow                1.10.0          mkl_py35heddcb22_0  
tensorflow-base           1.10.0          mkl_py35h3c3e929_0  
tensorflow-gpu            1.10.0               hf154084_0    anaconda
cudnn                     7.1.3                 cuda8.0_0  
cuda80                    1.0                           0    soumith
numpy                     1.15.2          py35_blas_openblashd3ea46f_0  [blas_openblas]  conda-forge
numpy-base                1.14.3           py35h2b20989_0  


pip3 install mtcnn


There are three different section of this project.

  1. Data pre-processing
  2. Training and testing
  3. Demo

We will go through the details in the following sections.

This repository is for 300W-LP, AFLW2000, and BIWI datasets.

1. Data pre-processing

[For lazy people just like me]

If you don't want to re-download every dataset images and do the pre-processing again, or maybe you don't even care about the data structure in the folder. Just download the file from the following link, and replace the data folder.

Google drive

Now you can skip to the "Training and testing" stage.


In the paper, we define Protocol 1 and Protocol 2.

# Protocol 1

Training: 300W-LP (A set of subsets: {AFW.npz, AFW_Flip.npz, HELEN.npz, HELEN_Flip.npz, IBUG.npz, IBUG_Flip.npz, LFPW.npz, LFPW_Flip.npz})
Testing: AFLW2000.npz or BIWI_noTrack.npz

# Protocol 2

Training: BIWI(70%)-> BIWI_train.npz
Testing: BIWI(30%)-> BIWI_test.npz

(Note that type1 (300W-LP, AFLW2000) datasets have the same image arrangement, and I categorize them as type1. It is not about Protocal 1 or 2.)

If you want to do the pre-processing from the beginning, you need to download the dataset first.

Download the datasets

Put 300W-LP and AFLW2000 folders under data/type1/, and put BIWI folder under data/

Run pre-processing

# For 300W-LP and AFLW2000 datasets

cd data/type1

# For BIWI dataset

cd data

2. Training and testing

# Training

# Testing
# Note that we calculate the MAE of yaw, pitch, roll independently, and average them into one single MAE for evaluation.

Just remember to check which model type you want to use in the shell script and you are good to go.

3. Demo

You need a webcam to correctly process the demo file.

Note the the center of the color axes is the detected face center. Ideally, each frame should have new face detection results. However, if the face detection fails, the previous detection results will be used to estimate poses.

LBP is fast enough for real-time face detection, while MTCNN is much more accurate but slow.

(2019/08/30 update!) SSD face detection is robust and fast! I borrow some face detector code from

# LBP face detector (fast but often miss detecting faces)
cd demo

# MTCNN face detector (slow but accurate)
cd demo

# SSD face detector (fast and accurate)
cd demo

4. Conversion to tensorflow frozen graph

cd training_and_testing
python --trained-model-dir-path ../pre-trained/300W_LP_models/fsanet_var_capsule_3_16_2_21_5 --output-dir-path <your_output_dir>

Above command will generate the tensorflow frozen graph in <your_output_dir>/converted-models/tf/fsanet_var_capsule_3_16_2_21_5.pb

Modules explanation:

  1. ssr_G_model:

  • Two-stream structure for extracting the features.
  1. ssr_feat_S_model:

  • Generating fine-grained structure mapping from different scoring functions.
  • Apply the mapping on to the features and generate primary capsules.
  1. ssr_aggregation_model:

  • Feed the primary capsules into capsule layer and output the final aggregated capsule features. And divide them into 3 stages.
  1. ssr_F_model:

  • Taking the previous 3 stages features for Soft-Stagewise Regression (SSR) module. Each stage further splits into three parts: prediction, dynamic index shifting, and dynamic scaling. This part please check the '[IJCAI18] SSR-Net' for more detail explanation.
  1. SSRLayer:

  • Taking the prediction, dynamic index shifting, and dynamic scaling for the final regression output. In this case, there are three outputs (yaw, pitch, roll).

3rd party implementation

Awesome VR flight simluation with FSANet ONNX format


[CVPR19] FSA-Net: Learning Fine-Grained Structure Aggregation for Head Pose Estimation from a Single Image





No releases published


No packages published