The MXNet Implementation of Stacked Hourglass and Stacked SAT for Robust 2D and 3D Face Alignment.

This repository contains several heatmap based approaches like stacked Hourglass and stacked Scale Aggregation Topology (SAT) for robust 2D and 3D face alignment. Some popular blocks such as bottleneck residual block, inception residual block, parallel and multi-scale (HPM) residual block and channel aggregation block (CAB) are also provided for building the topology of the deep face alignment network. All the codes in this repo are implemented in Python and MXNet.

The models for 2D face alignment are verified on IBUG, COFW and 300W test datasets by the normalised mean error (NME) respectively. For 3D face alignment, the 3D pre-trained models are compared on AFLW2000-3D with the most recent state-of-the-art methods.

The training/validation dataset and testset are in below table:

The training/validation dataset and testset are in below table:

Data Download Link Description BaiduCloud or GoogleDrive, 490M 2D training/validation dataset and IBUG, COFW, 300W testset BaiduCloud or GoogleDrive, 1.54G 3D training/validation dataset and AFLW2000-3D testset

The performances of 2D pre-trained models are shown below. Accuracy is reported as the Normalised Mean Error (NME). To facilitate comparison with other methods on these datasets, we give mean error normalised by the eye centre distance. Each training model is denoted by Topology^StackBlock (d = DownSampling Steps) - BlockType - OtherParameters.

Model Model Size IBUG COFW 300W Download Link
Hourglass2(d=4)-Resnet 26MB 7.719 6.776 6.482 BaiduCloud or GoogleDrive
Hourglass2(d=3)-HPM 38MB 7.249 6.378 6.049 BaiduCloud or GoogleDrive
Hourglass2(d=4)-CAB 46MB 7.168 6.123 5.684 BaiduCloud or GoogleDrive
SAT2(d=3)-CAB 40MB 7.052 5.999 5.618 BaiduCloud or GoogleDrive
Hourglass2(d=3)-CAB 37MB 6.974 5.983 5.647 BaiduCloud or GoogleDrive

The performances of 3D pre-trained models are shown below. Accuracy is reported as the Normalised Mean Error (NME). The mean error is normalised by the square root of the ground truth bounding box size.

Model Model Size AFLW2000-3D Download Link
SAT2(d=3)-CAB-3D 40MB 3.072 BaiduCloud or GoogleDrive
Hourglass2(d=3)-CAB-3D 37MB 3.005 BaiduCloud or GoogleDrive

Note: More pre-trained models will be added soon.


This repository has been tested under the following environment:

  • Python 2.7
  • Ubuntu 18.04
  • Mxnet-cu90 (==1.3.0)


  1. Prepare the environment.

  2. Clone the repository.

  3. Type make to build necessary cxx libs.


  • Download the training/validation dataset and unzip it to your project directory.

  • You can define different loss-type/network topology/dataset in

  • You can edit and run sh or use python to train your models. The following commands are some examples. Our experiments were conducted on a GTX 1080Ti GPU.

(1) Train stacked Scale Aggregation Topology (SAT) networks with channel aggregation block (CAB).

CUDA_VISIBLE_DEVICES='0' python --network satnet --prefix ./model/model-sat2d3-cab/model --per-batch-size 16 --lr 1e-4 --lr-epoch-step '20,35,45'

(2) Train stacked Hourglass models with parallel and multi-scale (HPM) residual block.

CUDA_VISIBLE_DEVICES='0' python --network hourglass --prefix ./model/model-hg2d3-hpm/model --per-batch-size 16 --lr 1e-4 --lr-epoch-step '20,35,45'


  • Download the ESSH model from BaiduCloud or GoogleDrive and place it in ./essh-model/.

  • Download the pre-trained model and place it in ./models/.

  • You can use python to test your models for 2D and 3D face alignment.


To evaluate pre-trained models on IBUG, COFW, 300W and AFLW2000-3D testset, you can use 'python' to obtain the Normalised Mean Error (NME) on the testset. We give some examples below.

  1. Evaluate model Hourglass2(d=3)-CAB with 2D landmarks on IBUG testset.
python --dataset ibug --prefix ./models/model-hg2d3-cab/model --epoch 0 --gpu 0 --landmark-type 2d
  1. Evaluate model SAT2(d=3)-CAB with 2D landmarks on COFW testset.
python --dataset cofw_testset --prefix ./models/model-sat2d3-cab/model --epoch 0 --gpu 0 --landmark-type 2d
  1. Evaluate model SAT2(d=3)-HPM with 2D landmarks on 300W testset.
python --dataset 300W --prefix ./models/model-hg2d3-hpm/model --epoch 0 --gpu 0 --landmark-type 2d
  1. Evaluate model Hourglass2(d=3)-CAB-3D with 3D landmarks on AFLW2000-3D testset.
python --dataset AFLW2000-3D --prefix ./models/model-hg2d3-cab-3d/model --epoch 0 --gpu 0 --landmark-type 3d


Results of 2D face alignment (inferenced from model Hourglass2(d=3)-CAB) are shown below.

Results on ALFW2000-3D dataset (inferenced from model Hourglass2(d=3)-CAB-3D) are shown below.