face.evoLVe: High-Performance Face Recognition Library based on PaddlePaddle & PyTorch
- Evolve to be more comprehensive, effective and efficient for face related analytics & applications! (WeChat News)
- About the name:
- "face" means this repo is dedicated for face related analytics & applications.
- "evolve" means unleash your greatness to be better and better. "LV" are capitalized to acknowledge the nurturing of Learning and Vision (LV) group, Nation University of Singapore (NUS).
- This work was done during Jian Zhao served as a short-term "Texpert" Research Scientist at Tencent FiT DeepSea AI Lab, Shenzhen, China.
Author | Jian Zhao |
---|---|
Homepage | https://zhaoj9014.github.io |
License
The code of face.evoLVe is released under the MIT License.
News
CLOSED 02 September 2021
: Baidu PaddlePaddle officially merged face.evoLVe to faciliate researches and applications on face-related analytics (Official Announcement).
CLOSED 03 July 2021
: Provides training code for the paddlepaddle framework.
CLOSED 04 July 2019
: We will share several publicly available datasets on face anti-spoofing/liveness detection to facilitate related research and analytics.
CLOSED 07 June 2019
: We are training a better-performing IR-152 model on MS-Celeb-1M_Align_112x112, and will release the model soon.
CLOSED 23 May 2019
: We share three publicly available datasets to facilitate research on heterogeneous face recognition and analytics. Please refer to Sec. Data Zoo for details.
CLOSED 23 Jan 2019
: We share the name lists and pair-wise overlapping lists of several widely-used face recognition datasets to help researchers/engineers quickly remove the overlapping parts between their own private datasets and the public datasets. Please refer to Sec. Data Zoo for details.
CLOSED 23 Jan 2019
: The current distributed training schema with multi-GPUs under PyTorch and other mainstream platforms parallels the backbone across multi-GPUs while relying on a single master to compute the final bottleneck (fully-connected/softmax) layer. This is not an issue for conventional face recognition with moderate number of identities. However, it struggles with large-scale face recognition, which requires recognizing millions of identities in the real world. The master can hardly hold the oversized final layer while the slaves still have redundant computation resource, leading to small-batch training or even failed training. To address this problem, we are developing a highly-elegant, effective and efficient distributed training schema with multi-GPUs under PyTorch, supporting not only the backbone, but also the head with the fully-connected (softmax) layer, to facilitate high-performance large-scale face recognition. We will added this support into our repo.
CLOSED 22 Jan 2019
: We have released two feature extraction APIs for extracting features from pre-trained models, implemented with PyTorch build-in functions and OpenCV, respectively. Please check ./util/extract_feature_v1.py
and ./util/extract_feature_v2.py
.
CLOSED 22 Jan 2019
: We are fine-tuning our released IR-50 model on our private Asia face data, which will be released soon to facilitate high-performance Asia face recognition.
CLOSED 21 Jan 2019
: We are training a better-performing IR-50 model on MS-Celeb-1M_Align_112x112, and will replace the current model soon.
Contents
- Introduction
- Pre-Requisites
- Usage
- Face Alignment
- Data Processing
- Training and Validation
- Data Zoo
- Model Zoo
- Achievement
- Acknowledgement
- Donation
- Citation
face.evoLVe for High-Performance Face Recognition
Introduction
- This repo provides a comprehensive face recognition library for face related analytics & applications, including face alignment (detection, landmark localization, affine transformation, etc.), data processing (e.g., augmentation, data balancing, normalization, etc.), various backbones (e.g., ResNet, IR, IR-SE, ResNeXt, SE-ResNeXt, DenseNet, LightCNN, MobileNet, ShuffleNet, DPN, etc.), various losses (e.g., Softmax, Focal, Center, SphereFace, CosFace, AmSoftmax, ArcFace, Triplet, etc.) and bags of tricks for improving performance (e.g., training refinements, model tweaks, knowledge distillation, etc.).
- The current distributed training schema with multi-GPUs under PyTorch and other mainstream platforms parallels the backbone across multi-GPUs while relying on a single master to compute the final bottleneck (fully-connected/softmax) layer. This is not an issue for conventional face recognition with moderate number of identities. However, it struggles with large-scale face recognition, which requires recognizing millions of identities in the real world. The master can hardly hold the oversized final layer while the slaves still have redundant computation resource, leading to small-batch training or even failed training. To address this problem, this repo provides a highly-elegant, effective and efficient distributed training schema with multi-GPUs under PyTorch, supporting not only the backbone, but also the head with the fully-connected (softmax) layer, to facilitate high-performance large-scale face recognition.
- All data before & after alignment, source codes and trained models are provided.
- This repo can help researchers/engineers develop high-performance deep face recognition models and algorithms quickly for practical use and deployment.
Pre-Requisites
- Linux or macOS
- Python 3.7 (for training & validation) and Python 2.7 (for visualization w/ tensorboardX)
- PyTorch 1.0 (for traininig & validation, install w/
pip install torch torchvision
) - MXNet 1.3.1 (optional, for data processing, install w/
pip install mxnet-cu90
) - TensorFlow 1.12 (optional, for visualization, install w/
pip install tensorflow-gpu
) - tensorboardX 1.6 (optional, for visualization, install w/
pip install tensorboardX
) - OpenCV 3.4.5 (install w/
pip install opencv-python
) - bcolz 1.2.0 (install w/
pip install bcolz
)
While not required, for optimal performance it is highly recommended to run the code using a CUDA enabled GPU. We used 4-8 NVIDIA Tesla P40 in parallel.
Usage
- Clone the repo:
git clone https://github.com/ZhaoJ9014/face.evoLVe.PyTorch.git
. mkdir data checkpoint log
at appropriate directory to store your train/val/test data, checkpoints and training logs.- Prepare your train/val/test data (refer to Sec. Data Zoo for publicly available face related databases), and ensure each database folder has the following structure:
./data/db_name/ -> id1/ -> 1.jpg -> ... -> id2/ -> 1.jpg -> ... -> ... -> ... -> ...
- Refer to the codes of corresponding sections for specific purposes.
Face Alignment
- This section is based on the work of MTCNN.
- Folder:
./align
- Face detection, landmark localization APIs and visualization toy example with ipython notebook:
from PIL import Image from detector import detect_faces from visualization_utils import show_results img = Image.open('some_img.jpg') # modify the image path to yours bounding_boxes, landmarks = detect_faces(img) # detect bboxes and landmarks for all faces in the image show_results(img, bounding_boxes, landmarks) # visualize the results
- Face alignment API (perform face detection, landmark localization and alignment with affine transformations on a whole database folder
source_root
with the directory structure as demonstrated in Sec. Usage, and store the aligned results to a new folderdest_root
with the same directory structure):python face_align.py -source_root [source_root] -dest_root [dest_root] -crop_size [crop_size] # python face_align.py -source_root './data/test' -dest_root './data/test_Aligned' -crop_size 112
- For macOS users, there is no need to worry about
*.DS_Store
files which may ruin your data, since they will be automatically removed when you run the scripts. - Keynotes for customed use: 1) specify the arguments of
source_root
,dest_root
andcrop_size
to your own values when you runface_align.py
; 2) pass your customedmin_face_size
,thresholds
andnms_thresholds
values to thedetect_faces
function ofdetector.py
to match your practical requirements; 3) if you find the speed using face alignment API is a bit slow, you can call face resize API to firstly resize the image whose smaller size is larger than a threshold (specify the arguments ofsource_root
,dest_root
andmin_side
to your own values) before calling the face alignment API:python face_resize.py
Data Processing
- Folder:
./balance
- Remove low-shot data API (remove the low-shot classes with less than
min_num
samples in the training setroot
with the directory structure as demonstrated in Sec. Usage for data balance and effective model training):python remove_lowshot.py -root [root] -min_num [min_num] # python remove_lowshot.py -root './data/train' -min_num 10
- Keynotes for customed use: specify the arguments of
root
andmin_num
to your own values when you runremove_lowshot.py
. - We prefer to include other data processing tricks, e.g., augmentation (flip horizontally, scale hue/satuation/brightness with coefficients uniformly drawn from [0.6,1.4], add PCA noise with a coefficient sampled from a normal distribution N(0,0.1), etc.), weighted random sampling, normalization, etc. to the main training script in Sec. Training and Validation to be self-contained.
Training and Validation
-
Folder:
./
-
Configuration API (configurate your overall settings for training & validation)
config.py
:import torch configurations = { 1: dict( SEED = 1337, # random seed for reproduce results DATA_ROOT = '/media/pc/6T/jasonjzhao/data/faces_emore', # the parent root where your train/val/test data are stored MODEL_ROOT = '/media/pc/6T/jasonjzhao/buffer/model', # the root to buffer your checkpoints LOG_ROOT = '/media/pc/6T/jasonjzhao/buffer/log', # the root to log your train/val status BACKBONE_RESUME_ROOT = './', # the root to resume training from a saved checkpoint HEAD_RESUME_ROOT = './', # the root to resume training from a saved checkpoint BACKBONE_NAME = 'IR_SE_50', # support: ['ResNet_50', 'ResNet_101', 'ResNet_152', 'IR_50', 'IR_101', 'IR_152', 'IR_SE_50', 'IR_SE_101', 'IR_SE_152'] HEAD_NAME = 'ArcFace', # support: ['Softmax', 'ArcFace', 'CosFace', 'SphereFace', 'Am_softmax'] LOSS_NAME = 'Focal', # support: ['Focal', 'Softmax'] INPUT_SIZE = [112, 112], # support: [112, 112] and [224, 224] RGB_MEAN = [0.5, 0.5, 0.5], # for normalize inputs to [-1, 1] RGB_STD = [0.5, 0.5, 0.5], EMBEDDING_SIZE = 512, # feature dimension BATCH_SIZE = 512, DROP_LAST = True, # whether drop the last batch to ensure consistent batch_norm statistics LR = 0.1, # initial LR NUM_EPOCH = 125, # total epoch number (use the firt 1/25 epochs to warm up) WEIGHT_DECAY = 5e-4, # do not apply to batch_norm parameters MOMENTUM = 0.9, STAGES = [35, 65, 95], # epoch stages to decay learning rate DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu"), MULTI_GPU = True, # flag to use multiple GPUs; if you choose to train with single GPU, you should first run "export CUDA_VISILE_DEVICES=device_id" to specify the GPU card you want to use GPU_ID = [0, 1, 2, 3], # specify your GPU ids PIN_MEMORY = True, NUM_WORKERS = 0, ), }
-
Train & validation API (all folks about training & validation, i.e., import package, hyperparameters & data loaders, model & loss & optimizer, train & validation & save checkpoint)
train.py
. Since MS-Celeb-1M serves as an ImageNet in the filed of face recognition, we pre-train the face.evoLVe models on MS-Celeb-1M and perform validation on LFW, CFP_FF, CFP_FP, AgeDB, CALFW, CPLFW and Vggface2_FP. Let's dive into details together step by step.- Import necessary packages:
import torch import torch.nn as nn import torch.optim as optim import torchvision.transforms as transforms import torchvision.datasets as datasets from config import configurations from backbone.model_resnet import ResNet_50, ResNet_101, ResNet_152 from backbone.model_irse import IR_50, IR_101, IR_152, IR_SE_50, IR_SE_101, IR_SE_152 from head.metrics import ArcFace, CosFace, SphereFace, Am_softmax from loss.focal import FocalLoss from util.utils import make_weights_for_balanced_classes, get_val_data, separate_irse_bn_paras, separate_resnet_bn_paras, warm_up_lr, schedule_lr, perform_val, get_time, buffer_val, AverageMeter, accuracy from tensorboardX import SummaryWriter from tqdm import tqdm import os
- Initialize hyperparameters:
cfg = configurations[1] SEED = cfg['SEED'] # random seed for reproduce results torch.manual_seed(SEED) DATA_ROOT = cfg['DATA_ROOT'] # the parent root where your train/val/test data are stored MODEL_ROOT = cfg['MODEL_ROOT'] # the root to buffer your checkpoints LOG_ROOT = cfg['LOG_ROOT'] # the root to log your train/val status BACKBONE_RESUME_ROOT = cfg['BACKBONE_RESUME_ROOT'] # the root to resume training from a saved checkpoint HEAD_RESUME_ROOT = cfg['HEAD_RESUME_ROOT'] # the root to resume training from a saved checkpoint BACKBONE_NAME = cfg['BACKBONE_NAME'] # support: ['ResNet_50', 'ResNet_101', 'ResNet_152', 'IR_50', 'IR_101', 'IR_152', 'IR_SE_50', 'IR_SE_101', 'IR_SE_152'] HEAD_NAME = cfg['HEAD_NAME'] # support: ['Softmax', 'ArcFace', 'CosFace', 'SphereFace', 'Am_softmax'] LOSS_NAME = cfg['LOSS_NAME'] # support: ['Focal', 'Softmax'] INPUT_SIZE = cfg['INPUT_SIZE'] RGB_MEAN = cfg['RGB_MEAN'] # for normalize inputs RGB_STD = cfg['RGB_STD'] EMBEDDING_SIZE = cfg['EMBEDDING_SIZE'] # feature dimension BATCH_SIZE = cfg['BATCH_SIZE'] DROP_LAST = cfg['DROP_LAST'] # whether drop the last batch to ensure consistent batch_norm statistics LR = cfg['LR'] # initial LR NUM_EPOCH = cfg['NUM_EPOCH'] WEIGHT_DECAY = cfg['WEIGHT_DECAY'] MOMENTUM = cfg['MOMENTUM'] STAGES = cfg['STAGES'] # epoch stages to decay learning rate DEVICE = cfg['DEVICE'] MULTI_GPU = cfg['MULTI_GPU'] # flag to use multiple GPUs GPU_ID = cfg['GPU_ID'] # specify your GPU ids PIN_MEMORY = cfg['PIN_MEMORY'] NUM_WORKERS = cfg['NUM_WORKERS'] print("=" * 60) print("Overall Configurations:") print(cfg) print("=" * 60) writer = SummaryWriter(LOG_ROOT) # writer for buffering intermedium results
- Train & validation data loaders:
train_transform = transforms.Compose([ # refer to https://pytorch.org/docs/stable/torchvision/transforms.html for more build-in online data augmentation transforms.Resize([int(128 * INPUT_SIZE[0] / 112), int(128 * INPUT_SIZE[0] / 112)]), # smaller side resized transforms.RandomCrop([INPUT_SIZE[0], INPUT_SIZE[1]]), transforms.RandomHorizontalFlip(), transforms.ToTensor(), transforms.Normalize(mean = RGB_MEAN, std = RGB_STD), ]) dataset_train = datasets.ImageFolder(os.path.join(DATA_ROOT, 'imgs'), train_transform) # create a weighted random sampler to process imbalanced data weights = make_weights_for_balanced_classes(dataset_train.imgs, len(dataset_train.classes)) weights = torch.DoubleTensor(weights) sampler = torch.utils.data.sampler.WeightedRandomSampler(weights, len(weights)) train_loader = torch.utils.data.DataLoader( dataset_train, batch_size = BATCH_SIZE, sampler = sampler, pin_memory = PIN_MEMORY, num_workers = NUM_WORKERS, drop_last = DROP_LAST ) NUM_CLASS = len(train_loader.dataset.classes) print("Number of Training Classes: {}".format(NUM_CLASS)) lfw, cfp_ff, cfp_fp, agedb, calfw, cplfw, vgg2_fp, lfw_issame, cfp_ff_issame, cfp_fp_issame, agedb_issame, calfw_issame, cplfw_issame, vgg2_fp_issame = get_val_data(DATA_ROOT)
- Define and initialize model (backbone & head):
BACKBONE_DICT = {'ResNet_50': ResNet_50(INPUT_SIZE), 'ResNet_101': ResNet_101(INPUT_SIZE), 'ResNet_152': ResNet_152(INPUT_SIZE), 'IR_50': IR_50(INPUT_SIZE), 'IR_101': IR_101(INPUT_SIZE), 'IR_152': IR_152(INPUT_SIZE), 'IR_SE_50': IR_SE_50(INPUT_SIZE), 'IR_SE_101': IR_SE_101(INPUT_SIZE), 'IR_SE_152': IR_SE_152(INPUT_SIZE)} BACKBONE = BACKBONE_DICT[BACKBONE_NAME] print("=" * 60) print(BACKBONE) print("{} Backbone Generated".format(BACKBONE_NAME)) print("=" * 60) HEAD_DICT = {'ArcFace': ArcFace(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS, device_id = GPU_ID), 'CosFace': CosFace(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS, device_id = GPU_ID), 'SphereFace': SphereFace(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS, device_id = GPU_ID), 'Am_softmax': Am_softmax(in_features = EMBEDDING_SIZE, out_features = NUM_CLASS, device_id = GPU_ID)} HEAD = HEAD_DICT[HEAD_NAME] print("=" * 60) print(HEAD) print("{} Head Generated".format(HEAD_NAME)) print("=" * 60)
- Define and initialize loss function:
LOSS_DICT = {'Focal': FocalLoss(), 'Softmax': nn.CrossEntropyLoss()} LOSS = LOSS_DICT[LOSS_NAME] print("=" * 60) print(LOSS) print("{} Loss Generated".format(LOSS_NAME)) print("=" * 60)
- Define and initialize optimizer:
if BACKBONE_NAME.find("IR") >= 0: backbone_paras_only_bn, backbone_paras_wo_bn = separate_irse_bn_paras(BACKBONE) # separate batch_norm parameters from others; do not do weight decay for batch_norm parameters to improve the generalizability _, head_paras_wo_bn = separate_irse_bn_paras(HEAD) else: backbone_paras_only_bn, backbone_paras_wo_bn = separate_resnet_bn_paras(BACKBONE) # separate batch_norm parameters from others; do not do weight decay for batch_norm parameters to improve the generalizability _, head_paras_wo_bn = separate_resnet_bn_paras(HEAD) OPTIMIZER = optim.SGD([{'params': backbone_paras_wo_bn + head_paras_wo_bn, 'weight_decay': WEIGHT_DECAY}, {'params': backbone_paras_only_bn}], lr = LR, momentum = MOMENTUM) print("=" * 60) print(OPTIMIZER) print("Optimizer Generated") print("=" * 60)
- Whether resume from a checkpoint or not:
if BACKBONE_RESUME_ROOT and HEAD_RESUME_ROOT: print("=" * 60) if os.path.isfile(BACKBONE_RESUME_ROOT) and os.path.isfile(HEAD_RESUME_ROOT): print("Loading Backbone Checkpoint '{}'".format(BACKBONE_RESUME_ROOT)) BACKBONE.load_state_dict(torch.load(BACKBONE_RESUME_ROOT)) print("Loading Head Checkpoint '{}'".format(HEAD_RESUME_ROOT)) HEAD.load_state_dict(torch.load(HEAD_RESUME_ROOT)) else: print("No Checkpoint Found at '{}' and '{}'. Please Have a Check or Continue to Train from Scratch".format(BACKBONE_RESUME_ROOT, HEAD_RESUME_ROOT)) print("=" * 60)
- Whether use multi-GPU or not:
if MULTI_GPU: # multi-GPU setting BACKBONE = nn.DataParallel(BACKBONE, device_ids = GPU_ID) BACKBONE = BACKBONE.to(DEVICE) else: # single-GPU setting BACKBONE = BACKBONE.to(DEVICE)
- Minor settings prior to training:
DISP_FREQ = len(train_loader) // 100 # frequency to display training loss & acc NUM_EPOCH_WARM_UP = NUM_EPOCH // 25 # use the first 1/25 epochs to warm up NUM_BATCH_WARM_UP = len(train_loader) * NUM_EPOCH_WARM_UP # use the first 1/25 epochs to warm up batch = 0 # batch index
- Training & validation & save checkpoint (use the first 1/25 epochs to warm up -- gradually increase LR to the initial value to ensure stable convergence):
for epoch in range(NUM_EPOCH): # start training process if epoch == STAGES[0]: # adjust LR for each training stage after warm up, you can also choose to adjust LR manually (with slight modification) once plaueau observed schedule_lr(OPTIMIZER) if epoch == STAGES[1]: schedule_lr(OPTIMIZER) if epoch == STAGES[2]: schedule_lr(OPTIMIZER) BACKBONE.train() # set to training mode HEAD.train() losses = AverageMeter() top1 = AverageMeter() top5 = AverageMeter() for inputs, labels in tqdm(iter(train_loader)): if (epoch + 1 <= NUM_EPOCH_WARM_UP) and (batch + 1 <= NUM_BATCH_WARM_UP): # adjust LR for each training batch during warm up warm_up_lr(batch + 1, NUM_BATCH_WARM_UP, LR, OPTIMIZER) # compute output inputs = inputs.to(DEVICE) labels = labels.to(DEVICE).long() features = BACKBONE(inputs) outputs = HEAD(features, labels) loss = LOSS(outputs, labels) # measure accuracy and record loss prec1, prec5 = accuracy(outputs.data, labels, topk = (1, 5)) losses.update(loss.data.item(), inputs.size(0)) top1.update(prec1.data.item(), inputs.size(0)) top5.update(prec5.data.item(), inputs.size(0)) # compute gradient and do SGD step OPTIMIZER.zero_grad() loss.backward() OPTIMIZER.step() # dispaly training loss & acc every DISP_FREQ if ((batch + 1) % DISP_FREQ == 0) and batch != 0: print("=" * 60) print('Epoch {}/{} Batch {}/{}\t' 'Training Loss {loss.val:.4f} ({loss.avg:.4f})\t' 'Training Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t' 'Training Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format( epoch + 1, NUM_EPOCH, batch + 1, len(train_loader) * NUM_EPOCH, loss = losses, top1 = top1, top5 = top5)) print("=" * 60) batch += 1 # batch index # training statistics per epoch (buffer for visualization) epoch_loss = losses.avg epoch_acc = top1.avg writer.add_scalar("Training_Loss", epoch_loss, epoch + 1) writer.add_scalar("Training_Accuracy", epoch_acc, epoch + 1) print("=" * 60) print('Epoch: {}/{}\t' 'Training Loss {loss.val:.4f} ({loss.avg:.4f})\t' 'Training Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t' 'Training Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format( epoch + 1, NUM_EPOCH, loss = losses, top1 = top1, top5 = top5)) print("=" * 60) # perform validation & save checkpoints per epoch # validation statistics per epoch (buffer for visualization) print("=" * 60) print("Perform Evaluation on LFW, CFP_FF, CFP_FP, AgeDB, CALFW, CPLFW and VGG2_FP, and Save Checkpoints...") accuracy_lfw, best_threshold_lfw, roc_curve_lfw = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, lfw, lfw_issame) buffer_val(writer, "LFW", accuracy_lfw, best_threshold_lfw, roc_curve_lfw, epoch + 1) accuracy_cfp_ff, best_threshold_cfp_ff, roc_curve_cfp_ff = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, cfp_ff, cfp_ff_issame) buffer_val(writer, "CFP_FF", accuracy_cfp_ff, best_threshold_cfp_ff, roc_curve_cfp_ff, epoch + 1) accuracy_cfp_fp, best_threshold_cfp_fp, roc_curve_cfp_fp = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, cfp_fp, cfp_fp_issame) buffer_val(writer, "CFP_FP", accuracy_cfp_fp, best_threshold_cfp_fp, roc_curve_cfp_fp, epoch + 1) accuracy_agedb, best_threshold_agedb, roc_curve_agedb = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, agedb, agedb_issame) buffer_val(writer, "AgeDB", accuracy_agedb, best_threshold_agedb, roc_curve_agedb, epoch + 1) accuracy_calfw, best_threshold_calfw, roc_curve_calfw = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, calfw, calfw_issame) buffer_val(writer, "CALFW", accuracy_calfw, best_threshold_calfw, roc_curve_calfw, epoch + 1) accuracy_cplfw, best_threshold_cplfw, roc_curve_cplfw = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, cplfw, cplfw_issame) buffer_val(writer, "CPLFW", accuracy_cplfw, best_threshold_cplfw, roc_curve_cplfw, epoch + 1) accuracy_vgg2_fp, best_threshold_vgg2_fp, roc_curve_vgg2_fp = perform_val(MULTI_GPU, DEVICE, EMBEDDING_SIZE, BATCH_SIZE, BACKBONE, vgg2_fp, vgg2_fp_issame) buffer_val(writer, "VGGFace2_FP", accuracy_vgg2_fp, best_threshold_vgg2_fp, roc_curve_vgg2_fp, epoch + 1) print("Epoch {}/{}, Evaluation: LFW Acc: {}, CFP_FF Acc: {}, CFP_FP Acc: {}, AgeDB Acc: {}, CALFW Acc: {}, CPLFW Acc: {}, VGG2_FP Acc: {}".format(epoch + 1, NUM_EPOCH, accuracy_lfw, accuracy_cfp_ff, accuracy_cfp_fp, accuracy_agedb, accuracy_calfw, accuracy_cplfw, accuracy_vgg2_fp)) print("=" * 60) # save checkpoints per epoch if MULTI_GPU: torch.save(BACKBONE.module.state_dict(), os.path.join(MODEL_ROOT, "Backbone_{}_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(BACKBONE_NAME, epoch + 1, batch, get_time()))) torch.save(HEAD.state_dict(), os.path.join(MODEL_ROOT, "Head_{}_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(HEAD_NAME, epoch + 1, batch, get_time()))) else: torch.save(BACKBONE.state_dict(), os.path.join(MODEL_ROOT, "Backbone_{}_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(BACKBONE_NAME, epoch + 1, batch, get_time()))) torch.save(HEAD.state_dict(), os.path.join(MODEL_ROOT, "Head_{}_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(HEAD_NAME, epoch + 1, batch, get_time())))
- Import necessary packages:
-
Now, you can start to play with face.evoLVe and run
train.py
. User friendly information will popped out on your terminal:-
About overall configuration:
-
About number of training classes:
-
About backbone details:
-
About head details:
-
About loss details:
-
About optimizer details:
-
About resume training:
-
About training status & statistics (when batch index reachs
DISP_FREQ
or at the end of each epoch): -
About validation statistics & save checkpoints (at the end of each epoch):
-
-
Monitor on-the-fly GPU occupancy with
watch -d -n 0.01 nvidia-smi
. -
Please refer to Sec. Model Zoo for specific model weights and corresponding performance.
-
Feature extraction API (extract features from pre-trained models)
./util/extract_feature_v1.py
(implemented with PyTorch build-in functions) and./util/extract_feature_v2.py
(implemented with OpenCV). -
Visualize training & validation statistics with tensorboardX (see Sec. Model Zoo):
tensorboard --logdir /media/pc/6T/jasonjzhao/buffer/log
Data Zoo
Database | Version | #Identity | #Image | #Frame | #Video | Download Link |
---|---|---|---|---|---|---|
LFW | Raw | 5,749 | 13,233 | - | - | Google Drive, Baidu Drive |
LFW | Align_250x250 | 5,749 | 13,233 | - | - | Google Drive, Baidu Drive |
LFW | Align_112x112 | 5,749 | 13,233 | - | - | Google Drive, Baidu Drive |
CALFW | Raw | 4,025 | 12,174 | - | - | Google Drive, Baidu Drive |
CALFW | Align_112x112 | 4,025 | 12,174 | - | - | Google Drive, Baidu Drive |
CPLFW | Raw | 3,884 | 11,652 | - | - | Google Drive, Baidu Drive |
CPLFW | Align_112x112 | 3,884 | 11,652 | - | - | Google Drive, Baidu Drive |
CASIA-WebFace | Raw_v1 | 10,575 | 494,414 | - | - | Baidu Drive |
CASIA-WebFace | Raw_v2 | 10,575 | 494,414 | - | - | Google Drive, Baidu Drive |
CASIA-WebFace | Clean | 10,575 | 455,594 | - | - | Google Drive, Baidu Drive |
MS-Celeb-1M | Clean | 100,000 | 5,084,127 | - | - | Google Drive |
MS-Celeb-1M | Align_112x112 | 85,742 | 5,822,653 | - | - | Google Drive |
Vggface2 | Clean | 8,631 | 3,086,894 | - | - | Google Drive |
Vggface2_FP | Align_112x112 | - | - | - | - | Google Drive, Baidu Drive |
AgeDB | Raw | 570 | 16,488 | - | - | Google Drive, Baidu Drive |
AgeDB | Align_112x112 | 570 | 16,488 | - | - | Google Drive, Baidu Drive |
IJB-A | Clean | 500 | 5,396 | 20,369 | 2,085 | Google Drive, Baidu Drive |
IJB-B | Raw | 1,845 | 21,798 | 55,026 | 7,011 | Google Drive |
CFP | Raw | 500 | 7,000 | - | - | Google Drive, Baidu Drive |
CFP | Align_112x112 | 500 | 7,000 | - | - | Google Drive, Baidu Drive |
Umdfaces | Align_112x112 | 8,277 | 367,888 | - | - | Google Drive, Baidu Drive |
CelebA | Raw | 10,177 | 202,599 | - | - | Google Drive, Baidu Drive |
CACD-VS | Raw | 2,000 | 163,446 | - | - | Google Drive, Baidu Drive |
YTF | Align_344x344 | 1,595 | - | 3,425 | 621,127 | Google Drive, Baidu Drive |
DeepGlint | Align_112x112 | 180,855 | 6,753,545 | - | - | Google Drive |
UTKFace | Align_200x200 | - | 23,708 | - | - | Google Drive, Baidu Drive |
BUAA-VisNir | Align_287x287 | 150 | 5,952 | - | - | Baidu Drive, PW: xmbc |
CASIA NIR-VIS 2.0 | Align_128x128 | 725 | 17,580 | - | - | Baidu Drive, PW: 883b |
Oulu-CASIA | Raw | 80 | 65,000 | - | - | Baidu Drive, PW: xxp5 |
NUAA-ImposterDB | Raw | 15 | 12,614 | - | - | Baidu Drive, PW: if3n |
CASIA-SURF | Raw | 1,000 | - | - | 21,000 | Baidu Drive, PW: izb3 |
CASIA-FASD | Raw | 50 | - | - | 600 | Baidu Drive, PW: h5un |
CASIA-MFSD | Raw | 50 | - | - | 600 | |
Replay-Attack | Raw | 50 | - | - | 1,200 | |
WebFace260M | Raw | 24M | 2M | - | https://www.face-benchmark.org/ |
- Remark: unzip CASIA-WebFace clean version with
unzip casia-maxpy-clean.zip cd casia-maxpy-clean zip -F CASIA-maxpy-clean.zip --out CASIA-maxpy-clean_fix.zip unzip CASIA-maxpy-clean_fix.zip
- Remark: after unzip, get image data & pair ground truths from AgeDB, CFP, LFW and VGGFace2_FP align_112x112 versions with
import numpy as np import bcolz import os def get_pair(root, name): carray = bcolz.carray(rootdir = os.path.join(root, name), mode='r') issame = np.load('{}/{}_list.npy'.format(root, name)) return carray, issame def get_data(data_root): agedb_30, agedb_30_issame = get_pair(data_root, 'agedb_30') cfp_fp, cfp_fp_issame = get_pair(data_root, 'cfp_fp') lfw, lfw_issame = get_pair(data_root, 'lfw') vgg2_fp, vgg2_fp_issame = get_pair(data_root, 'vgg2_fp') return agedb_30, cfp_fp, lfw, vgg2_fp, agedb_30_issame, cfp_fp_issame, lfw_issame, vgg2_fp_issame agedb_30, cfp_fp, lfw, vgg2_fp, agedb_30_issame, cfp_fp_issame, lfw_issame, vgg2_fp_issame = get_data(DATA_ROOT)
- Remark: We share
MS-Celeb-1M_Top1M_MID2Name.tsv
(Google Drive, Baidu Drive),VGGface2_ID2Name.csv
(Google Drive, Baidu Drive),VGGface2_FaceScrub_Overlap.txt
(Google Drive, Baidu Drive),VGGface2_LFW_Overlap.txt
(Google Drive, Baidu Drive),CASIA-WebFace_ID2Name.txt
(Google Drive, Baidu Drive),CASIA-WebFace_FaceScrub_Overlap.txt
(Google Drive, Baidu Drive),CASIA-WebFace_LFW_Overlap.txt
(Google Drive, Baidu Drive),FaceScrub_Name.txt
(Google Drive, Baidu Drive),LFW_Name.txt
(Google Drive, Baidu Drive),LFW_Log.txt
(Google Drive, Baidu Drive) to help researchers/engineers quickly remove the overlapping parts between their own private datasets and the public datasets. - Due to release license issue, for other face related databases, please make contact with us in person for more details.
Model Zoo
-
Model
Backbone Head Loss Training Data Download Link IR-50 ArcFace Focal MS-Celeb-1M_Align_112x112 Google Drive, Baidu Drive -
Setting
INPUT_SIZE: [112, 112]; RGB_MEAN: [0.5, 0.5, 0.5]; RGB_STD: [0.5, 0.5, 0.5]; BATCH_SIZE: 512 (drop the last batch to ensure consistent batch_norm statistics); Initial LR: 0.1; NUM_EPOCH: 120; WEIGHT_DECAY: 5e-4 (do not apply to batch_norm parameters); MOMENTUM: 0.9; STAGES: [30, 60, 90]; Augmentation: Random Crop + Horizontal Flip; Imbalanced Data Processing: Weighted Random Sampling; Solver: SGD; GPUs: 4 NVIDIA Tesla P40 in Parallel
-
Training & validation statistics
-
Performance
LFW CFP_FF CFP_FP AgeDB CALFW CPLFW Vggface2_FP 99.78 99.69 98.14 97.53 95.87 92.45 95.22
-
-
Model
Backbone Head Loss Training Data Download Link IR-50 ArcFace Focal Private Asia Face Data Google Drive, Baidu Drive -
Setting
INPUT_SIZE: [112, 112]; RGB_MEAN: [0.5, 0.5, 0.5]; RGB_STD: [0.5, 0.5, 0.5]; BATCH_SIZE: 1024 (drop the last batch to ensure consistent batch_norm statistics); Initial LR: 0.01 (initialize weights from the above model pre-trained on MS-Celeb-1M_Align_112x112); NUM_EPOCH: 80; WEIGHT_DECAY: 5e-4 (do not apply to batch_norm parameters); MOMENTUM: 0.9; STAGES: [20, 40, 60]; Augmentation: Random Crop + Horizontal Flip; Imbalanced Data Processing: Weighted Random Sampling; Solver: SGD; GPUs: 8 NVIDIA Tesla P40 in Parallel
-
Performance (please perform evaluation on your own Asia face benchmark dataset)
-
-
Model
Backbone Head Loss Training Data Download Link IR-152 ArcFace Focal MS-Celeb-1M_Align_112x112 Baidu Drive, PW: b197 -
Setting
INPUT_SIZE: [112, 112]; RGB_MEAN: [0.5, 0.5, 0.5]; RGB_STD: [0.5, 0.5, 0.5]; BATCH_SIZE: 256 (drop the last batch to ensure consistent batch_norm statistics); Initial LR: 0.01; NUM_EPOCH: 120; WEIGHT_DECAY: 5e-4 (do not apply to batch_norm parameters); MOMENTUM: 0.9; STAGES: [30, 60, 90]; Augmentation: Random Crop + Horizontal Flip; Imbalanced Data Processing: Weighted Random Sampling; Solver: SGD; GPUs: 4 NVIDIA Geforce RTX 2080 Ti in Parallel
-
Training & validation statistics
-
Performance
LFW CFP_FF CFP_FP AgeDB CALFW CPLFW Vggface2_FP 99.82 99.83 98.37 98.07 96.03 93.05 95.50
-
Achievement
-
2017 No.1 on ICCV 2017 MS-Celeb-1M Large-Scale Face Recognition Hard Set/Random Set/Low-Shot Learning Challenges. WeChat News, NUS ECE News, NUS ECE Poster, Award Certificate for Track-1, Award Certificate for Track-2, Award Ceremony.
-
2017 No.1 on National Institute of Standards and Technology (NIST) IARPA Janus Benchmark A (IJB-A) Unconstrained Face Verification challenge and Identification challenge. WeChat News.
-
State-of-the-art performance on
- MS-Celeb-1M (Challenge1 Hard Set Coverage@P=0.95: 79.10%; Challenge1 Random Set Coverage@P=0.95: 87.50%; Challenge2 Development Set Coverage@P=0.99: 100.00%; Challenge2 Base Set Top 1 Accuracy: 99.74%; Challenge2 Novel Set Coverage@P=0.99: 99.01%).
- IJB-A (1:1 Veification TAR@FAR=0.1: 99.6%±0.1%; 1:1 Veification TAR@FAR=0.01: 99.1%±0.2%; 1:1 Veification TAR@FAR=0.001: 97.9%±0.4%; 1:N Identification FNIR@FPIR=0.1: 1.3%±0.3%; 1:N Identification FNIR@FPIR=0.01: 5.4%±4.7%; 1:N Identification Rank1 Accuracy: 99.2%±0.1%; 1:N Identification Rank5 Accuracy: 99.7%±0.1%; 1:N Identification Rank10 Accuracy: 99.8%±0.1%).
- IJB-C (1:1 Veification TAR@FAR=1e-5: 82.6%).
- Labeled Faces in the Wild (LFW) (Accuracy: 99.85%±0.217%).
- Celebrities in Frontal-Profile (CFP) (Frontal-Profile Accuracy: 96.01%±0.84%; Frontal-Profile EER: 4.43%±1.04%; Frontal-Profile AUC: 99.00%±0.35%; Frontal-Frontal Accuracy: 99.64%±0.25%; Frontal-Frontal EER: 0.54%±0.37%; Frontal-Frontal AUC: 99.98%±0.03%).
- CMU Multi-PIE (Rank1 Accuracy Setting-1 under ±90°: 76.12%; Rank1 Accuracy Setting-2 under ±90°: 86.73%).
- MORPH Album2 (Rank1 Accuracy Setting-1: 99.65%; Rank1 Accuracy Setting-2: 99.26%).
- CACD-VS (Accuracy: 99.76%).
- FG-NET (Rank1 Accuracy: 93.20%).
Acknowledgement
- This repo is inspired by InsightFace.MXNet, InsightFace.PyTorch, ArcFace.PyTorch, MTCNN.MXNet and PretrainedModels.PyTorch.
- The work of Jian Zhao was partially supported by China Scholarship Council (CSC) grant 201503170248.
- We would like to thank Prof. Jiashi Feng, Dr. Jianshu Li, Mr. Yu Cheng (Learning and Vision group, National University of Singapore), Mr. Yuan Xin, Mr. Di Wu, Mr. Zhenyuan Shen, Mr. Jianwei Liu (Tencent FiT DeepSea AI Lab, China), Prof. Ran He, Prof. Junliang Xing, Mr. Xiang Wu (Institute of Automation, Chinese Academy of Sciences), Prof. Guosheng Hu (AnyVision Inc., U.K.), Dr. Lin Xiong (JD Digits, U.S.), Miss Yi Cheng (Panasonic R&D Center, Singapore) for helpful discussions.
Donation
-
Your donation is highly welcomed to help us further develop face.evoLVe to better facilitate more cutting-edge researches and applications on facial analytics and human-centric multi-media understanding. The donation QR code via Wechat is as below. Appreciate it very much
❤️
Citation
-
Please consult and consider citing the following papers:
@article{wang2021face, title={Face. evoLVe: A High-Performance Face Recognition Library}, author={Wang, Qingzhong and Zhang, Pengfei and Xiong, Haoyi and Zhao, Jian}, journal={arXiv preprint arXiv:2107.08621}, year={2021} } @article{tu2021image, title={Image-to-Video Generation via 3D Facial Dynamics}, author={Tu, Xiaoguang and Zou, Yingtian and Zhao, Jian and Ai, Wenjie and Dong, Jian and Yao, Yuan and Wang, Zhikang and Guo, Guodong and Li, Zhifeng and Liu, Wei and others}, journal={T-CSVT}, year={2021} } @article{tu2021joint, title={Joint Face Image Restoration and Frontalization for Recognition}, author={Tu, Xiaoguang and Zhao, Jian and Liu, Qiankun and Ai, Wenjie and Guo, Guodong and Li, Zhifeng and Liu, Wei and Feng, Jiashi}, journal={T-CSVT}, year={2021} } @article{zhao2020towards, title={Towards age-invariant face recognition}, author={Zhao, Jian and Yan, Shuicheng and Feng, Jiashi}, journal={T-PAMI}, year={2020} } @article{liang2020fine, title={Fine-grained facial expression recognition in the wild}, author={Liang, Liqian and Lang, Congyan and Li, Yidong and Feng, Songhe and Zhao, Jian}, journal={T-IFS}, pages={482--494}, year={2020} } @article{tu2020learning, title={Learning generalizable and identity-discriminative representations for face anti-spoofing}, author={Tu, Xiaoguang and Ma, Zheng and Zhao, Jian and Du, Guodong and Xie, Mei and Feng, Jiashi}, journal={T-IST}, pages={1--19}, year={2020} } @article{tu20203d, title={3D face reconstruction from a single image assisted by 2D face images in the wild}, author={Tu, Xiaoguang and Zhao, Jian and Xie, Mei and Jiang, Zihang and Balamurugan, Akshaya and Luo, Yao and Zhao, Yang and He, Lingxiao and Ma, Zheng and Feng, Jiashi}, journal={T-MM}, year={2020} } @inproceedings{wang2020learning, title={Learning to Detect Head Movement in Unconstrained Remote Gaze Estimation in the Wild}, author={Wang, Zhecan and Zhao, Jian and Lu, Cheng and Yang, Fan and Huang, Han and Guo, Yandong and others}, booktitle={WACV}, pages={3443--3452}, year={2020} } @article{zhao2019recognizing, title={Recognizing Profile Faces by Imagining Frontal View}, author={Zhao, Jian and Xing, Junliang and Xiong, Lin and Yan, Shuicheng and Feng, Jiashi}, journal={IJCV}, pages={1--19}, year={2019} } @article{kong2019cross, title={Cross-Resolution Face Recognition via Prior-Aided Face Hallucination and Residual Knowledge Distillation}, author={Kong, Hanyang and Zhao, Jian and Tu, Xiaoguang and Xing, Junliang and Shen, Shengmei and Feng, Jiashi}, journal={arXiv preprint arXiv:1905.10777}, year={2019} } @article{tu2019joint, title={Joint 3D face reconstruction and dense face alignment from a single image with 2D-assisted self-supervised learning}, author={Tu, Xiaoguang and Zhao, Jian and Jiang, Zihang and Luo, Yao and Xie, Mei and Zhao, Yang and He, Linxiao and Ma, Zheng and Feng, Jiashi}, journal={arXiv preprint arXiv:1903.09359}, year={2019} } @inproceedings{zhao2019multi, title={Multi-Prototype Networks for Unconstrained Set-based Face Recognition}, author={Zhao, Jian and Li, Jianshu and Tu, Xiaoguang and Zhao, Fang and Xin, Yuan and Xing, Junliang and Liu, Hengzhu and Yan, Shuicheng and Feng, Jiashi}, booktitle={IJCAI}, year={2019} } @inproceedings{zhao2019look, title={Look Across Elapse: Disentangled Representation Learning and Photorealistic Cross-Age Face Synthesis for Age-Invariant Face Recognition}, author={Zhao, Jian and Cheng, Yu and Cheng, Yi and Yang, Yang and Lan, Haochong and Zhao, Fang and Xiong, Lin and Xu, Yan and Li, Jianshu and Pranata, Sugiri and others}, booktitle={AAAI}, year={2019} } @article{tu2019joint, title={Joint 3D Face Reconstruction and Dense Face Alignment from A Single Image with 2D-Assisted Self-Supervised Learning}, author={Tu, Xiaoguang and Zhao, Jian and Jiang, Zihang and Luo, Yao and Xie, Mei and Zhao, Yang and He, Linxiao and Ma, Zheng and Feng, Jiashi}, journal={arXiv preprint arXiv:1903.09359}, year={2019} } @article{tu2019learning, title={Learning Generalizable and Identity-Discriminative Representations for Face Anti-Spoofing}, author={Tu, Xiaoguang and Zhao, Jian and Xie, Mei and Du, Guodong and Zhang, Hengsheng and Li, Jianshu and Ma, Zheng and Feng, Jiashi}, journal={arXiv preprint arXiv:1901.05602}, year={2019} } @article{zhao20183d, title={3D-Aided Dual-Agent GANs for Unconstrained Face Recognition}, author={Zhao, Jian and Xiong, Lin and Li, Jianshu and Xing, Junliang and Yan, Shuicheng and Feng, Jiashi}, journal={T-PAMI}, year={2018} } @inproceedings{zhao2018towards, title={Towards Pose Invariant Face Recognition in the Wild}, author={Zhao, Jian and Cheng, Yu and Xu, Yan and Xiong, Lin and Li, Jianshu and Zhao, Fang and Jayashree, Karlekar and Pranata, Sugiri and Shen, Shengmei and Xing, Junliang and others}, booktitle={CVPR}, pages={2207--2216}, year={2018} } @inproceedings{zhao3d, title={3D-Aided Deep Pose-Invariant Face Recognition}, author={Zhao, Jian and Xiong, Lin and Cheng, Yu and Cheng, Yi and Li, Jianshu and Zhou, Li and Xu, Yan and Karlekar, Jayashree and Pranata, Sugiri and Shen, Shengmei and others}, booktitle={IJCAI}, pages={1184--1190}, year={2018} } @inproceedings{zhao2018dynamic, title={Dynamic Conditional Networks for Few-Shot Learning}, author={Zhao, Fang and Zhao, Jian and Yan, Shuicheng and Feng, Jiashi}, booktitle={ECCV}, pages={19--35}, year={2018} } @inproceedings{zhao2017dual, title={Dual-agent gans for photorealistic and identity preserving profile face synthesis}, author={Zhao, Jian and Xiong, Lin and Jayashree, Panasonic Karlekar and Li, Jianshu and Zhao, Fang and Wang, Zhecan and Pranata, Panasonic Sugiri and Shen, Panasonic Shengmei and Yan, Shuicheng and Feng, Jiashi}, booktitle={NeurIPS}, pages={66--76}, year={2017} } @inproceedings{zhao122017marginalized, title={Marginalized cnn: Learning deep invariant representations}, author={Zhao12, Jian and Li, Jianshu and Zhao, Fang and Yan13, Shuicheng and Feng, Jiashi}, booktitle={BMVC}, year={2017} } @inproceedings{cheng2017know, title={Know you at one glance: A compact vector representation for low-shot learning}, author={Cheng, Yu and Zhao, Jian and Wang, Zhecan and Xu, Yan and Jayashree, Karlekar and Shen, Shengmei and Feng, Jiashi}, booktitle={ICCVW}, pages={1924--1932}, year={2017} } @inproceedings{xu2017high, title={High performance large scale face recognition with multi-cognition softmax and feature retrieval}, author={Xu, Yan and Cheng, Yu and Zhao, Jian and Wang, Zhecan and Xiong, Lin and Jayashree, Karlekar and Tamura, Hajime and Kagaya, Tomoyuki and Shen, Shengmei and Pranata, Sugiri and others}, booktitle={ICCVW}, pages={1898--1906}, year={2017} } @inproceedings{wangconditional, title={Conditional Dual-Agent GANs for Photorealistic and Annotation Preserving Image Synthesis}, author={Wang, Zhecan and Zhao, Jian and Cheng, Yu and Xiao, Shengtao and Li, Jianshu and Zhao, Fang and Feng, Jiashi and Kassim, Ashraf}, booktitle={BMVCW}, } @inproceedings{li2017integrated, title={Integrated face analytics networks through cross-dataset hybrid training}, author={Li, Jianshu and Xiao, Shengtao and Zhao, Fang and Zhao, Jian and Li, Jianan and Feng, Jiashi and Yan, Shuicheng and Sim, Terence}, booktitle={ACM MM}, pages={1531--1539}, year={2017} } @article{xiong2017good, title={A good practice towards top performance of face recognition: Transferred deep feature fusion}, author={Xiong, Lin and Karlekar, Jayashree and Zhao, Jian and Cheng, Yi and Xu, Yan and Feng, Jiashi and Pranata, Sugiri and Shen, Shengmei}, journal={arXiv preprint arXiv:1704.00438}, year={2017} } @article{zhao2017robust, title={Robust lstm-autoencoders for face de-occlusion in the wild}, author={Zhao, Fang and Feng, Jiashi and Zhao, Jian and Yang, Wenhan and Yan, Shuicheng}, journal={T-IP}, volume={27}, number={2}, pages={778--790}, year={2017} } @inproceedings{li2016robust, title={Robust face recognition with deep multi-view representation learning}, author={Li, Jianshu and Zhao, Jian and Zhao, Fang and Liu, Hao and Li, Jing and Shen, Shengmei and Feng, Jiashi and Sim, Terence}, booktitle={ACM MM}, pages={1068--1072}, year={2016} }