SipMask

This is the official implementation of "SipMask: Spatial Information Preservation for Fast Image and Video Instance Segmentation (ECCV2020)" built on the open-source mmdetection and maskrcnn-benchmark.

Single-stage method for both image and video instance segmentation.
Two different versions are provided: high-accuracy version and real-time (fast) version.
Image instance segmentation is built on both mmdetection and maskrcnn-benchmark.
Video instance segmentation is built on mmdetection.
Datasets: MS COCO for image instance segmentation and YouTube-VIS for video instance segmentation.
Json results of SipMask+ResNet50 on YouTube-VIS: Single-scale training (32.5) and Multi-scale training (33.7)

Introduction

Single-stage instance segmentation approaches have recently gained popularity due to their speed and simplicity, but are still lagging behind in accuracy, compared to two-stage methods. We propose a fast single-stage instance segmentation method, called SipMask, that preserves instance-specific spatial information by separating the mask prediction of an instance to different sub-regions of a detected bounding-box. Our main contribution is a novel light-weight spatial preservation (SP) module that generates a separate set of spatial coefficients for each sub-region within a bounding-box, leading to improved mask predictions. It also enables accurate delineation of spatially adjacent instances. Further, we introduce a mask alignment weighting loss and a feature alignment scheme to better correlate mask prediction with object detection.

SipMask-benchmark (image instance segmentation)

This project is built on the official implementation of FCOS, which is based on maskrcnn-benchmark.
High-quality version is provided.
Please use SipMask-benchmark and refer to INSTALL.md for installation.
PyTorch1.1.0 and cuda9.0/10.0 are used by me.

Train with multiple GPUs

python -m torch.distributed.launch --nproc_per_node=4 --master_port=$((RANDOM+10000)) tools/train_net.py --config-file ${CONFIG_FILE} DATALOADER.NUM_WORKERS 2 OUTPUT_DIR ${OUTPUT_PATH}
e.g.,
python -m torch.distributed.launch --nproc_per_node=4 --master_port=$((RANDOM+10000)) tools/train_net.py --config-file configs/sipmask/sipmask_R_50_FPN_1x.yaml DATALOADER.NUM_WORKERS 2 OUTPUT_DIR training_dir/sipmask_R_50_FPN_1x

Test with a single GPU

python tools/test_net.py --config-file ${CONFIG_FILE} MODEL.WEIGHT ${CHECKPOINT_FILE} TEST.IMS_PER_BATCH 4
e.g.,
python tools/test_net.py --config-file configs/sipmask/sipmask_R_50_FPN_1x.yaml MODEL.WEIGHT  training_dir/SipMask_R50_1x.pth TEST.IMS_PER_BATCH 4

CONFIG_FILE of Sipmask is under the folder of SipMask-benchmark/configs/sipmask.

Results

name	backbone	input size	epoch	ms-train	val. box AP	val. mask AP	download
SipMask	R50	800 × 1333	1x	no	39.5	34.2	model
SipMask	R101	800 × 1333	3x	yes	44.1	37.8	model

SipMask-mmdetection (image instance segmentation)

This project is built on mmdetection.
High-quality version and real-time version are both provided.
Please use SipMask-mmdetection and refer to INSTALL.md for installation.
PyTorch1.1.0, cuda9.0/10.0, and mmcv0.4.3 are used by me.

Train with multiple GPUs

./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM} [optional arguments]
e.g.,
CUDA_VISIBLE_DEVICES=0,1,2,3 ./tools/dist_train.sh configs/sipmask/sipmask_r50_caffe_fpn_gn_1x_4gpu.py 4 --validate

Test with a single GPU

python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--out ${RESULT_FILE}] [--eval ${EVAL_METRICS}] [--show]
e.g., 
python tools/test.py ./configs/sipmask/sipmask_r50_caffe_fpn_gn_1x_4gpu.py ./work_dirs/sipmask_r50_caffe_1x.pth --out results.pkl --eval bbox segm

Inference with saved results

With our trained model, detection results of an image can be visualized using the following command.

python ./demo/sipmask_demo.py ${CONFIG_FILE} ${CHECKPOINT_FILE} ${IMAGE_FILE} [--out ${OUT_PATH}]
e.g.,
python ./demo/sipmask_demo.py ./configs/sipmask/sipmask_r50_caffe_fpn_gn_1x_4gpu.py ./sipmask_r50_caffe_1x.pth ./demo/demo.jpg --out ./demo/aa.jpg

CONFIG_FILE of Sipmask is under the folder of SipMask-mmdetection/configs/sipmask.

Results

name	backbone	input size	epoch	ms-train	GN	val. box AP	val. mask AP	download
SipMask	R50	800×1333	1x	no	yes	38.2	33.5	model
SipMask	R50	800×1333	2x	yes	yes	40.8	35.6	model
SipMask	R101	800×1333	4x	yes	yes	43.6	37.8	model
SipMask	R50	544×544	6x	yes	no	36.0	31.7	model
SipMask	R50	544×544	10x	yes	yes	37.1	32.4	model
SipMask	R101	544×544	6x	yes	no	38.4	33.6	model
SipMask	R101	544×544	10x	yes	yes	40.3	34.8	model
SipMask++	R101-D	544×544	6x	yes	no	40.1	35.2	model
SipMask++	R101-D	544×544	10x	yes	yes	41.3	36.1	model

GN indicates group normalization used in prediction branch.
Model with the input size of 800×1333 fcoses on high accuracy, which is trained in RetinaNet style.
Model with the input size of 544×544 fcoses on fast speed, which is trained in SSD style.
++ indicates adding deformable convolutions with interval of 3 in backbone and mask re-scoring module.

SipMask-VIS (video instance segmentation)

This project is an implementation for video instance segmenation based on mmdetection.
Please use SipMask-VIS and refer to INSTALL.md for installation.
PyTorch1.1.0, cuda9.0/10.0, and mmcv0.2.12 are used by me.

Please note that, to run YouTube-VIS dataset like MaskTrackRCNN, install the cocoapi for youtube-vis instead of installing the original cocoapi for coco as follows.

pip install git+https://github.com/youtubevos/cocoapi.git#"egg=pycocotools&subdirectory=PythonAPI"
or
cd SipMask-VIS/pycocotools/cocoapi/PythonAPI
python setup.py build_ext install

Train with multiple GPUs

./tools/dist_train.sh ${CONFIG_FILE} ${GPU_NUM}
e.g.,
CUDA_VISIBLE_DEVICES=0,1,2,3 ./toools/dist_train.sh ./configs/sipmask/sipmask_r50_caffe_fpn_gn_1x_4gpu.py 4

Test with a single GPU

python tools/test_video.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--out ${RESULT_FILE}] --eval segm
e.g.,
python ./tools/test_video.py configs/sipmask/sipmask_r50_caffe_fpn_gn_1x_4gpu.py ./work_dirs/sipmask_r50_fpn_1x.pth --out results.pkl --eval segm

If you want to save the results of video instance segmentation, please use the following command:

python tools/test_video.py ${CONFIG_FILE} ${CHECKPOINT_FILE} [--out ${RESULT_FILE}] --eval segm --show --save_path= ${SAVE_PATH}

CONFIG_FILE of SipMask-VIS is under the folder of SipMask-VIS/configs/sipmask.
The model pretrained on MS COCO dataset is used for weight initialization.

Results

name	backbone	input size	epoch	ms-train	val. mask AP	download
SipMask	R50	360 × 640	1x	no	32.5	model
SipMask	R50	360 × 640	1x	yes	33.7	model

The generated results on YouTube-VIS should be uploaded to codalab for evaluation.

Citation

If the project helps your research, please cite this paper.

@article{Cao_SipMask_ECCV_2020,
  author =       {Jiale Cao and Rao Muhammad Anwer and Hisham Cholakkal and Fahad Shahbaz Khan and Yanwei Pang and Ling Shao},
  title =        {SipMask: Spatial Information Preservation for Fast Image and Video Instance Segmentation},
  journal =      {Proc. European Conference on Computer Vision},
  year =         {2020}
}

Acknowledgement

Many thanks to the open source codes, i.e., FCOS, mmdetection, YOLACT, and MaskTrack RCNN.

Borda / SipMask

SipMask

Introduction

SipMask-benchmark (image instance segmentation)

Train with multiple GPUs

Test with a single GPU

Results

SipMask-mmdetection (image instance segmentation)

Train with multiple GPUs

Test with a single GPU

Inference with saved results

Results

SipMask-VIS (video instance segmentation)

Train with multiple GPUs

Test with a single GPU

Results

Citation

Acknowledgement

About

Languages