EmbedMask: Embedding Coupling for One-stage Instance Segmentation

This is repository is for the paper:

EmbedMask: Embedding Coupling for One-stage Instance Segmentation;
Hui Ying, Zhaojin Huang, Shu Liu, Tianjia Shao and Kun Zhou;
arXiv preprint arXiv:1912.01954

The full paper is available at: https://arxiv.org/abs/1912.01954. But some optimizations have been applied to the codes now, beyond the paper.

Installation

This EmbedMask implementation is based on FCOS, which is also based on maskrcnn-benchmark. Therefore the installation is the same. Please check INSTALL.md of FCOS for installation instructions.

Data Preparation

We do training and inference in the COCO dataset. If you want perform training and inference as well, please download the dataset, and put the data in the right place like follows.

mkdir -p datasets/coco
ln -s /path_to_coco_dataset/annotations datasets/coco/annotations
ln -s /path_to_coco_dataset/train2014 datasets/coco/train2014
ln -s /path_to_coco_dataset/test2014 datasets/coco/test2014
ln -s /path_to_coco_dataset/val2014 datasets/coco/val2014

Pretrained Models

The pretrained models can be downloaded from here. And you should place them in the 'models' directory.

Here are the scores of these models on COCO 'test2014' set:

Model	AP
R50_1x	34.7
R101_1x	36.4
R101_ms_3x	38.1

Demo

Once you have finished the installation and downloaded the pretrained models, you can run a quick demo by the following instructions which use the settings and from 'embed_mask_R50_1x'.

# assume that you are under the root directory of this project,
mkdir -p demo/ouput
python demo/embed_mask_demo.py \
    --config-file configs/embed_mask/embed_mask_R50_1x.yaml \
    --weights models/embed_mask_R50_1x.pth

Inference

The following inference command line run inference on coco minival split:

python tools/test_net.py \
    --config-file configs/embed_mask/embed_mask_R50_1x.yaml \
    MODEL.WEIGHT models/embed_mask_R50_1x.pth \
    TEST.IMS_PER_BATCH 4

Speed Testing

The following inference command line run speed testing on coco minival split:

python tools/test_net.py \
    --config-file configs/embed_mask/embed_mask_R50_1x.yaml \
    --speed_only \
    MODEL.WEIGHT models/embed_mask_R50_1x.pth \
    TEST.IMS_PER_BATCH 1 \
    MODEL.EMBED_MASK.POSTPROCESS_MASKS True

Training

The following command line will train 'embed_mask_R50_1x' on 4 GPUs with batchsize 16:

python -m torch.distributed.launch \
    --nproc_per_node=4 \
    --master_port=$((RANDOM + 10000)) \
    tools/train_net.py \
    --config-file configs/embed_mask/embed_mask_R50_1x.yaml \
    DATALOADER.NUM_WORKERS 4 \
    SOLVER.IMS_PER_BATCH 16 \
    OUTPUT_DIR training_dir/embed_mask_R50_1x

Citations

Please consider citing our paper in your publications if the project helps your research. BibTeX reference is as follows.

@misc{ying2019embedmask,
    title={EmbedMask: Embedding Coupling for One-stage Instance Segmentation},
    author={Hui Ying and Zhaojin Huang and Shu Liu and Tianjia Shao and Kun Zhou},
    year={2019},
    eprint={1912.01954},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Davidnet / EmbedMask