hukenovs / easyportrait

EasyPortrait - Face Parsing and Portrait Segmentation Dataset

EasyPortrait - Face Parsing and Portrait Segmentation Dataset

We introduce a large-scale image dataset EasyPortrait for portrait segmentation and face parsing. Proposed dataset can be used in several tasks, such as background removal in conference applications, teeth whitening, face skin enhancement, red eye removal or eye colorization, and so on.

EasyPortrait dataset size is about 91.78GB, and it contains 40,000 RGB images (~38.3K FullHD images) with high quality annotated masks. This dataset is divided into training set, validation set and test set by subject user_id. The training set includes 30,000 images, the validation set includes 4,000 images, and the test set includes 6,000 images.

For more information see our paper EasyPortrait – Face Parsing and Portrait Segmentation Dataset.

🔥 Changelog

• 2023/11/13: We release EasyPortrait 2.0. ✌️
  • 40,000 RGB images (~38.3K FullHD images)
  • Added diversity by region, race, human emotions and lighting conditions
  • The data was further cleared and new ones were added
  • Train/val/test split: (30,000) 75% / (4,000) 10% / (6,000) 15% by subject user_id
  • Multi-gpu training and testing
  • Added new models for face parsing and portrait segmentation
  • Dataset size is 91.78GB
  • 13,705 unique persons
• 2023/02/23: EasyPortrait (Initial Dataset) 💪
  • Dataset size is 26GB
  • 20,000 RGB images (~17.5K FullHD images) with 9 classes annotated
  • Train/val/test split: (14,000) 70% / (2,000) 10% / (4,000) 20% by subject user_id
  • 8,377 unique persons

Old EasyPortrait dataset is also available into branch EasyPortrait_v1!

Downloads

Link	Size
images	91.8 GB
annotations	657.1 MB
meta	1.9 MB
train set	68.3 GB
validation set	10.7 GB
test set	12.8 GB

Also, you can download EasyPortrait dataset from Kaggle.

Structure

.
├── images.zip
│   ├── train/         # Train set: 30k
│   ├── val/           # Validation set: 4k
│   ├── test/          # Test set: 6k
├── annotations.zip
│   ├── train/     
│   ├── val/       
│   ├── test/      
├── meta.zip       # Meta-information (width, height, brightness, imhash, user_id)
...

Models

We provide some pre-trained models as the baseline for portrait segmentation and face parsing. We use mean Intersection over Union (mIoU) as the main metric.

Portrait segmentation:

Model Name	Parameters (M)	Input shape	mIoU
BiSeNet-V2	56.5	384 x 384	97.95
DANet	190.2	384 x 384	98.63
DeepLabv3	260	384 x 384	98.63
ExtremeC3Net	0.15	384 x 384	96.54
Fast SCNN	6.13	384 x 384	97.64
FCN + MobileNetv2	31.17	384 x 384	98.19
FPN + ResNet50	108.91	1024 × 1024	98.54
FPN + ResNet50	108.91	512 × 512	98.64
FPN + ResNet50	108.91	384 x 384	98.64
FPN + ResNet50	108.91	224 × 224	98.31
SegFormer-B0	14.9	1024 × 1024	98.74
SegFormer-B0	14.9	512 × 512	98.66
SegFormer-B0	14.9	384 x 384	98.61
SegFormer-B0	14.9	224 × 224	98.17
SINet	0.13	384 x 384	93.32

Face parsing:

Model Name	Parameters (M)	Input shape	mIoU
BiSeNet-V2	56.5	384 x 384	76.72
DANet	190.2	384 x 384	79.3
DeepLabv3	260	384 x 384	79.11
EHANet	44.81	384 x 384	72.56
Fast SCNN	6.13	384 x 384	67.56
FCN + MobileNetv2	31.17	384 x 384	75.23
FPN + ResNet50	108.91	1024 × 1024	85.37
FPN + ResNet50	108.91	512 × 512	83.33
FPN + ResNet50	108.91	384 x 384	81.83
FPN + ResNet50	108.91	224 × 224	75.6
SegFormer-B0	14.9	1024 × 1024	85.42
SegFormer-B0	14.9	512 × 512	83.19
SegFormer-B0	14.9	384 x 384	81.38
SegFormer-B0	14.9	224 × 224	74.83

Annotations

Annotations are presented as 2D-arrays, images in *.png format with several classes:

Index	Class
0	BACKGROUND
1	PERSON
2	SKIN
3	LEFT_BROW
4	RIGHT_BROW
5	LEFT_EYE
6	RIGHT_EYE
7	LIPS
8	TEETH

Also, we provide some additional meta-information for dataset in annotations/meta.zip file:

	image_name	user_id	height	width	set	brightness
0	a753e021-...	56...	720	960	train	126
1	4ff04492-...	ba...	1920	1440	test	173
2	e8934c99-...	1d...	1920	1440	val	187

where:

• image_name - image file name without extension
• user_id - unique anonymized user ID
• height - image height
• width - image width
• brightness - image brightness
• set - "train", "test" or "val" for train / test / val subsets respectively

Images

Training, Evaluation and Testing on EasyPortrait

The code is based on MMSegmentation with 0.30.0 version.

Models were trained and evaluated on 8 NVIDIA V100 GPUs with CUDA 11.2.

For installation process follow the instructions here and use the requirements.txt file in our repository.

Training

For single GPU mode:

python ./pipelines/tools/train.py ./pipelines/local_configs/easy_portrait_experiments/<model_dir>/<config_file>.py --gpu-id <GPU_ID>

For distributed training mode:

./pipelines/tools/dist_train.sh ./pipelines/local_configs/easy_portrait_experiments/<model_dir>/<config_file>.py <NUM_GPUS>

Evaluation

For single GPU mode:

python ./pipelines/tools/test.py <PATH_TO_MODEL_CONFIG>  <PATH_TO_CHECKPOINT> --gpu-id <GPU_ID> --eval mIoU

For distributed evaluation mode:

./pipelines/tools/dist_test.sh <PATH_TO_MODEL_CONFIG>  <PATH_TO_CHECKPOINT> <NUM_GPUS> --eval mIoU

Run demo

python ./pipelines/demo/image_demo.py <PATH_TO_IMG> <PATH_TO_MODEL_CONFIG> <PATH_TO_CHECKPOINT> --palette=easy_portrait --out-file=<PATH_TO_OUT_FILE>

Authors and Credits

• Alexander Kapitanov
• Karina Kvanchiani
• Elizaveta Petrova
• Karen Efremyan
• Alexander Sautin

Links

• arXiv
• Paperswithcode
• Kaggle
• Habr
• Gitlab

Citation

You can cite the paper using the following BibTeX entry:

@article{EasyPortrait,
    title={EasyPortrait - Face Parsing and Portrait Segmentation Dataset},
    author={Kapitanov, Alexander and Kvanchiani, Karina and Kirillova Sofia},
    journal={arXiv preprint arXiv:2304.13509},
    year={2023}
}

License

This work is licensed under a variant of Creative Commons Attribution-ShareAlike 4.0 International License.

Please see the specific license.