CenterMask

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

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

*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 (