CenterMask2

[CenterMask(original code)][vovnet-detectron2][arxiv] [BibTeX]

CenterMask2 is an upgraded implementation on top of detectron2 beyond original CenterMask based on maskrcnn-benchmark.

CenterMask : Real-Time Anchor-Free Instance Segmentation (CVPR 2020)
Youngwan Lee and Jongyoul Park
Electronics and Telecommunications Research Institute (ETRI)
pre-print : https://arxiv.org/abs/1911.06667

Highlights

First anchor-free one-stage instance segmentation. To the best of our knowledge, CenterMask is the first instance segmentation on top of anchor-free object detection (15/11/2019).
Toward Real-Time: CenterMask-Lite. This works provide not only large-scale CenterMask but also lightweight CenterMask-Lite that can run at real-time speed (> 30 fps).
State-of-the-art performance. CenterMask outperforms Mask R-CNN, TensorMask, and ShapeMask at much faster speed and CenterMask-Lite models also surpass YOLACT or YOLACT++ by large margins.
Well balanced (speed/accuracy) backbone network, VoVNetV2. VoVNetV2 shows better performance and faster speed than ResNe(X)t or HRNet.

Updates

CenterMask2 has been released. (20/02/2020)
Lightweight VoVNet has ben released. (26/02/2020)
Panoptic-CenterMask has been released. (31/03/2020)
code update for compatibility with pytorch1.7 and the latest detectron2 (22/12/2020)
Results on COCO val

Note

We measure the inference time of all models with batch size 1 on the same V100 GPU machine.

pytorch1.7.0
CUDA 10.1
cuDNN 7.3
multi-scale augmentation
Unless speficified, no Test-Time Augmentation (TTA)

CenterMask

Method	Backbone	lr sched	inference time	mask AP	box AP	download
Mask R-CNN (detectron2)	R-50	3x	0.055	37.2	41.0	model \| metrics
Mask R-CNN (detectron2)	V2-39	3x	0.052	39.3	43.8	model \| metrics
CenterMask (maskrcnn-benchmark)	V2-39	3x	0.070	38.5	43.5	link
CenterMask2	V2-39	3x	0.050	39.7	44.2	model \| metrics
Mask R-CNN (detectron2)	R-101	3x	0.070	38.6	42.9	model \| metrics
Mask R-CNN (detectron2)	V2-57	3x	0.058	39.7	44.2	model \| metrics
CenterMask (maskrcnn-benchmark)	V2-57	3x	0.076	39.4	44.6	link
CenterMask2	V2-57	3x	0.058	40.5	45.1	model \| metrics
Mask R-CNN (detectron2)	X-101	3x	0.129	39.5	44.3	model \| metrics
Mask R-CNN (detectron2)	V2-99	3x	0.076	40.3	44.9	model \| metrics
CenterMask (maskrcnn-benchmark)	V2-99	3x	0.106	40.2	45.6	link
CenterMask2	V2-99	3x	0.077	41.4	46.0	model \| metrics
CenterMask2 (TTA)	V2-99	3x	-	42.5	48.6	model \| metrics

TTA denotes Test-Time Augmentation (multi-scale test).

CenterMask-Lite

Method	Backbone	lr sched	inference time	mask AP	box AP	download
YOLACT550	R-50	4x	0.023	28.2	30.3	link
CenterMask (maskrcnn-benchmark)	V-19	4x	0.023	32.4	35.9	link
CenterMask2-Lite	V-19	4x	0.023	32.8	35.9	model \| metrics
YOLACT550	R-101	4x	0.030	28.2	30.3	link
YOLACT550++	R-50	4x	0.029	34.1	-	link
YOLACT550++	R-101	4x	0.036	34.6	-	link
CenterMask (maskrcnn-benchmark)	V-39	4x	0.027	36.3	40.7	link
CenterMask2-Lite	V-39	4x	0.028	36.7	40.9	model \| metrics

Note that The inference time is measured on Titan Xp GPU for fair comparison with YOLACT.

Lightweight VoVNet backbone

Method	Backbone	Param.	lr sched	inference time	mask AP	box AP	download
CenterMask2-Lite	MobileNetV2	3.5M	4x	0.021	27.2	29.8	model \| metrics
CenterMask2-Lite	V-19	11.2M	4x	0.023	32.8	35.9	model \| metrics
CenterMask2-Lite	V-19-Slim	3.1M	4x	0.021	29.8	32.5	model \| metrics
CenterMask2-Lite	V-19Slim-DW	1.8M	4x	0.020	27.1	29.5	model \| metrics

DW and Slim denote depthwise separable convolution and a thiner model with half the channel size, respectively.
Params. means the number of parameters of backbone.

Deformable VoVNet Backbone

Method	Backbone	lr sched	inference time	mask AP	box AP	download
CenterMask2	V2-39	3x	0.050	39.7	44.2	model \| metrics
CenterMask2	V2-39-DCN	3x	0.061	40.3	45.1	model \| metrics
CenterMask2	V2-57	3x	0.058	40.5	45.1	model \| metrics
CenterMask2	V2-57-DCN	3x	0.071	40.9	45.5	model \| metrics
CenterMask2	V2-99	3x	0.077	41.4	46.0	model \| metrics
CenterMask2	V2-99-DCN	3x	0.110	42.0	46.9	model \| metrics

DCN denotes deformable convolutional networks v2. Note that we apply deformable convolutions from stage 3 to 5 in backbones.

Panoptic-CenterMask

Method	Backbone	lr sched	inference time	mask AP	box AP	PQ	download
Panoptic-FPN	R-50	3x	0.063	40.0	36.5	41.5	model \| metrics
Panoptic-CenterMask	R-50	3x	0.063	41.4	37.3	42.0	model \| metrics
Panoptic-FPN	V-39	3x	0.063	42.8	38.5	43.4	model \| metrics
Panoptic-CenterMask	V-39	3x	0.066	43.4	39.0	43.7	model \| metrics
Panoptic-FPN	R-101	3x	0.078	42.4	38.5	43.0	model \| metrics
Panoptic-CenterMask	R-101	3x	0.076	43.5	39.0	43.6	model \| metrics
Panoptic-FPN	V-57	3x	0.070	43.4	39.2	44.3	model \| metrics
Panoptic-CenterMask	V-57	3x	0.071	43.9	39.6	44.5	model \| metrics
Panoptic-CenterMask	V-99	3x	0.091	45.1	40.6	45.4	model \| metrics

Installation

All you need to use centermask2 is detectron2. It's easy!
you just install detectron2 following INSTALL.md.
Prepare for coco dataset following this instruction.

Training

ImageNet Pretrained Models

We provide backbone weights pretrained on ImageNet-1k dataset for detectron2.

To train a model, run

cd centermask2
python train_net.py --config-file "configs/<config.yaml>"

For example, to launch CenterMask training with VoVNetV2-39 backbone on 8 GPUs, one should execute:

cd centermask2
python train_net.py --config-file "configs/centermask/centermask_V_39_eSE_FPN_ms_3x.yaml" --num-gpus 8

Evaluation

Model evaluation can be done similarly:

if you want to inference with 1 batch --num-gpus 1
--eval-only
MODEL.WEIGHTS path/to/the/model.pth

cd centermask2
wget https://dl.dropbox.com/s/tczecsdxt10uai5/centermask2-V-39-eSE-FPN-ms-3x.pth
python train_net.py --config-file "configs/centermask/centermask_V_39_eSE_FPN_ms_3x.yaml" --num-gpus 1 --eval-only MODEL.WEIGHTS centermask2-V-39-eSE-FPN-ms-3x.pth

TODO

[x] Adding Lightweight models
[ ] Applying CenterMask for PointRend or Panoptic-FPN.

Citing CenterMask

If you use VoVNet, please use the following BibTeX entry.

@inproceedings{lee2019energy,
  title = {An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection},
  author = {Lee, Youngwan and Hwang, Joong-won and Lee, Sangrok and Bae, Yuseok and Park, Jongyoul},
  booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops},
  year = {2019}
}

@inproceedings{lee2020centermask,
  title={CenterMask: Real-Time Anchor-Free Instance Segmentation},
  author={Lee, Youngwan and Park, Jongyoul},
  booktitle={CVPR},
  year={2020}
}

Special Thanks to

mask scoring for detectron2 by Sangrok Lee
FCOS_for_detectron2 by AdeliDet team.

youngwanLEE / centermask2

readme