CenterMask : Real-Time Anchor-Free Instance Segmentation

Abstract

We propose a simple yet efficient anchor-free instance segmentation, called CenterMask, that adds a novel spatial attention-guided mask (SAG-Mask) branch to anchor-free one stage object detector (FCOS) in the same vein with Mask R-CNN. Plugged into the FCOS object detector, the SAG-Mask branch predicts a segmentation mask on each box with the spatial attention map that helps to focus on informative pixels and suppress noise. We also present an improved VoVNetV2 backbone networks with two effective strategies: (1) residual connection for alleviating the saturation problem of larger VoVNet and (2) effective Squeeze-Excitation (eSE) dealing with the information loss problem of original SE. With SAG-Mask and VoVNetV2, we deign CenterMask and CenterMask-Lite that are targeted to large and small models, respectively. CenterMask outperforms all previous state-of-the-art models at a much faster speed. CenterMask-Lite also achieves 33.4% mask AP / 38.0% box AP, outperforming YOLACT by 2.6 / 7.0 AP gain, respectively, at over 35fps on Titan Xp. We hope that CenterMask and VoVNetV2 can serve as a solid baseline of real-time instance segmentation and backbone network for various vision tasks, respectively.

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

• Open the official repo and code will be released after refactoring. (05/12/2019)
• Release code and MobileNetV2 & ResNet backbone models shown in the [paper]. (10/12/2019)
• Upload the VoVNetV2 backbone models. (02/01/2020)
• Open VoVNetV2 backbone for Detectron2 --> vovnet-detectron2. (08/01/2020)
• Upload CenterMask-Lite models trained for 48 epochs outperforming YOLACT or YOLACT++. (14/01/2020)
• centermask2 has been released. (20/02/2020)

Models

Environment

• V100 or Titan Xp GPU
• CUDA 10.0
• cuDNN7.3
• pytorch1.1
• Implemented on fcos and maskrcn-benchmark
• GoogleDrive weight download

coco test-dev results

Detector	Backbone	epoch	Mask AP (AP/APs/APm/APl)	Box AP (AP/APs/APm/APl)	Time (ms)	GPU	Weight
ShapeMask	R-101-FPN	N/A	37.4/16.1/40.1/53.8	42.2/24.9/45.2/52.7	125	V100	-
TensorMask	R-101-FPN	72	37.1/17.4/39.1/51.6	-	380	V100	-
RetinaMask	R-101-FPN	24	34.7/14.3/36.7/50.5	41.4/23.0/44.5/53.0	98	V100	-
Mask R-CNN	R-101-FPN	24	37.9/18.1/40.3/53.3	42.2/24.9/45.2/52.7	94	V100	-
CenterMask	R-101-FPN	24	38.3/17.7/40.8/54.5	43.1/25.2/46.1/54.4	72	V100	link
CenterMask	X-101-FPN	36	39.6/19.7/42.0/55.2	44.6/27.1/47.2/55.2	123	V100	link
CenterMask	V2-99-FPN	36	40.6/20.1/42.8/57.0	45.8/27.8/48.3/57.6	84	V100	link
YOLACT-400	R-101-FPN	48	24.9/5.0/25.3/45.0	28.4/10.7/28.9/43.1	22	Xp	-
CenterMask-Lite	MV2-FPN	48	26.7/9.0/27.0/40.9	30.2/14.2/31.9/40.9	20	Xp	link
YOLACT-550	R-50-FPN	48	28.2/9.2/29.3/44.8	30.3/14.0/31.2/43.0	23	Xp	-
CenterMask-Lite	V2-19-FPN	48	32.4/13.6/33.8/47.2	35.9/19.6/38.0/45.9	23	Xp	link
YOLACT-550	R-101-FPN	48	29.8/9.9/31.3/47.7	31.0/14.4/31.8/43.7	30	Xp	-
YOLACT-550++	R-50-FPN	48	34.1/11.7/36.1/53.6	-	29	Xp	-
YOLACT-550++	R-101-FPN	48	34.6/11.9/36.8/55.1	-	36	Xp	-
CenterMask-Lite	R-50-FPN	48	32.9/12.9/34.7/48.7	36.7/18.7/39.4/48.2	29	Xp	link
CenterMask-Lite	V2-39-FPN	48	36.3/15.6/38.1/53.1	40.7/22.4/43.2/53.5	28	Xp	link

Note that RetinaMask, Mask R-CNN, and CenterMask are implemented by using same baseline code(maskrcnn-benchmark) and all models are trained using multi-scale training augmentation.
We expect that if we implement our CenterMask based on detectron2, it will get better performance.
24/36/48/72 epoch are same as 2x/3x/4x/6x training schedule in detectron, respectively.
Training CenterMask-Lite models longer (24 --> 48 epochs same as YOLACT) boosts ther performance, widening the performance gap from YOLACT and even YOLACT++.

coco val2017 results

Detector	Backbone	epoch	Mask AP (AP/APs/APm/APl)	Box AP (AP/APs/APm/APl)	Time (ms)	Weight
CenterMask	MV2-FPN	36	31.2/14.5/32.8/46.3	35.5/20.6/38.0/46.8	56	link
CenterMask	V2-19-FPN	36	34.7/17.3/37.5/49.6	39.7/24.6/42.7/50.8	59	link
Mask R-CNN	R-50-FPN	24	35.9/17.1/38.9/52.0	39.7/24.0/43.0/50.8	77	link
CenterMask	R-50-FPN	24	36.4/17.3/39.5/52.7	41.2/24.9/45.1/53.0	72	link
CenterMask	V2-39-FPN	24	37.7/17.9/40.8/54.3	42.6/25.3/46.3/55.2	70	link
Mask R-CNN	R-50-FPN	36	36.5/17.9/39.2/52.5	40.5/24.7/43.7/52.2	77	link
CenterMask	R-50-FPN	36	37.0/17.6/39.7/53.8	41.7/24.8/45.1/54.5	72	link
CenterMask	V2-39-FPN	36	38.5/19.0/41.5/54.7	43.5/27.1/46.9/55.9	70	link
Mask R-CNN	R-101-FPN	24	37.8/18.5/40.7/54.9	42.2/25.8/45.8/54.0	94	link
CenterMask	R-101-FPN	24	38.0/18.2/41.3/55.2	43.1/25.7/47.0/55.6	91	link
CenterMask	V2-57-FPN	24	38.5/18.6/41.9/56.2	43.8/26.7/47.4/57.1	76	link
Mask R-CNN	R-101-FPN	36	38.0/18.4/40.8/55.2	42.4/25.4/45.5/55.2	94	link
CenterMask	R-101-FPN	36	38.6/19.2/42.0/56.1	43.7/27.2/47.6/56.7	91	link
CenterMask	V2-57-FPN	36	39.4/19.6/42.9/55.9	44.6/27.7/48.3/57.3	76	link
Mask R-CNN	X-101-32x8d-FPN	24	38.9/19.6/41.6/55.7	43.7/27.6/46.9/55.9	165	link
CenterMask	X-101-32x8d-FPN	24	39.1/19.6/42.5/56.1	44.3/26.9/48.5/57.0	157	link
CenterMask	V2-99-FPN	24	39.6/19.6/43.1/56.9	44.8/27.6/49.0/57.7	106	link
Mask R-CNN	X-101-32x8d-FPN	36	38.6/19.7/41.1/55.2	43.6/27.3/46.7/55.6	165	link
CenterMask	X-101-32x8d-FPN	36	39.1/18.5/42.3/56.4	44.4/26.7/47.7/57.1	157	link
CenterMask	V2-99-FPN	36	40.2/20.6/43.5/57.3	45.6/29.2/49.3/58.8	106	link

Note that the all models are trained using train-time augmentation (multi-scale).
The inference time of all models is measured on Titan Xp GPU.
24/36 epoch are same as x2/x3 training schedule in detectron, respectively.

Installation

Check INSTALL.md for installation instructions which is orginate from maskrcnn-benchmark.

Training

Follow the instructions of maskrcnn-benchmark guides.

If you want multi-gpu (e.g.,8) training,

export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/train_net.py --config-file "configs/centermask/centermask_R_50_FPN_1x.yaml" 

Evaluation

Follow the instruction of maskrcnn-benchmark

First of all, you have to download the weight file you want to inference.

For examaple (CenterMask-Lite-R-50),

multi-gpu evaluation & test batch size 16,

wget https://www.dropbox.com/s/2enqxenccz4xy6l/centermask-lite-R-50-ms-bs32-1x.pth
export NGPUS=8
python -m torch.distributed.launch --nproc_per_node=$NGPUS tools/test_net.py --config-file "configs/centermask/centermask_R_50_FPN_lite_res600_ms_bs32_1x.yaml"   TEST.IMS_PER_BATCH 16 MODEL.WEIGHT centermask-lite-R-50-ms-bs32-1x.pth

For single-gpu evaluation & test batch size 1,

wget https://www.dropbox.com/s/2enqxenccz4xy6l/centermask-lite-R-50-ms-bs32-1x.pth
CUDA_VISIBLE_DEVICES=0
python tools/test_net.py --config-file "configs/centermask/centermask_R_50_FPN_lite_res600_ms_bs32_1x.yaml" TEST.IMS_PER_BATCH 1 MODEL.WEIGHT centermask-lite-R-50-ms-bs32-1x.pth

TODO

• train-time augmentation + 3x schedule for comparing with detectron2 models
• ResNet-50 & ResNeXt-101-32x8d
• VoVNetV2 backbones
• VoVNetV2 backbones for Detectron2
• CenterMask in Detectron2
• quick-demo
• arxiv paper update

Performance

Citing CenterMask

Please cite our paper in your publications if it helps your research:

@article{lee2019centermask,
  title={CenterMask: Real-Time Anchor-Free Instance Segmentation},
  author={Lee, Youngwan and Park, Jongyoul},
  booktitle={CVPR},
  year={2020}
}