• Highlights
• Updates
• Installation
• Models

本项目nonzero算子Mindspore暂时无法支持，以及SOLOv2动态学习卷积核的思想暂时与Mindspore的静态图哲学理念暂不兼容。

head,mask_head,backbone,fpn,dateset已经完成。

This project hosts the code for implementing the SOLO algorithms for instance segmentation.

SOLO: Segmenting Objects by Locations,
Xinlong Wang, Tao Kong, Chunhua Shen, Yuning Jiang, Lei Li
In: Proc. European Conference on Computer Vision (ECCV), 2020
arXiv preprint (arXiv 1912.04488)

SOLOv2: Dynamic and Fast Instance Segmentation,
Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, Chunhua Shen
In: Proc. Advances in Neural Information Processing Systems (NeurIPS), 2020
arXiv preprint (arXiv 2003.10152)

Highlights

• Totally box-free: SOLO is totally box-free thus not being restricted by (anchor) box locations and scales, and naturally benefits from the inherent advantages of FCNs.
• Direct instance segmentation: Our method takes an image as input, directly outputs instance masks and corresponding class probabilities, in a fully convolutional, box-free and grouping-free paradigm.
• High-quality mask prediction: SOLOv2 is able to predict fine and detailed masks, especially at object boundaries.
• State-of-the-art performance: Our best single model based on ResNet-101 and deformable convolutions achieves 41.7% in AP on COCO test-dev (without multi-scale testing). A light-weight version of SOLOv2 executes at 31.3 FPS on a single V100 GPU and yields 37.1% AP.

Updates

• SOLOv2 implemented on detectron2 is released at adet. (07/12/20)
• Training speeds up (~1.7x faster) for all models. (03/12/20)
• SOLOv2 is available. Code and trained models of SOLOv2 are released. (08/07/2020)
• Light-weight models and R101-based models are available. (31/03/2020)
• SOLOv1 is available. Code and trained models of SOLO and Decoupled SOLO are released. (28/03/2020)

Installation

This implementation is based on mmdetection(v1.0.0). Please refer to INSTALL.md for installation and dataset preparation.

Models

For your convenience, we provide the following trained models on COCO (more models are coming soon).
If you need the models in PaddlePaddle framework, please refer to paddlepaddle/README.md.

Model	Multi-scale training	Testing time / im	AP (minival)	Link
SOLO_R50_1x	No	77ms	32.9	download
SOLO_R50_3x	Yes	77ms	35.8	download
SOLO_R101_3x	Yes	86ms	37.1	download
Decoupled_SOLO_R50_1x	No	85ms	33.9	download
Decoupled_SOLO_R50_3x	Yes	85ms	36.4	download
Decoupled_SOLO_R101_3x	Yes	92ms	37.9	download
SOLOv2_R50_1x	No	54ms	34.8	download
SOLOv2_R50_3x	Yes	54ms	37.5	download
SOLOv2_R101_3x	Yes	66ms	39.1	download
SOLOv2_R101_DCN_3x	Yes	97ms	41.4	download
SOLOv2_X101_DCN_3x	Yes	169ms	42.4	download

Light-weight models:

Model	Multi-scale training	Testing time / im	AP (minival)	Link
Decoupled_SOLO_Light_R50_3x	Yes	29ms	33.0	download
Decoupled_SOLO_Light_DCN_R50_3x	Yes	36ms	35.0	download
SOLOv2_Light_448_R18_3x	Yes	19ms	29.6	download
SOLOv2_Light_448_R34_3x	Yes	20ms	32.0	download
SOLOv2_Light_448_R50_3x	Yes	24ms	33.7	download
SOLOv2_Light_512_DCN_R50_3x	Yes	34ms	36.4	download

Disclaimer:

• Light-weight means light-weight backbone, head and smaller input size. Please refer to the corresponding config files for details.
• This is a reimplementation and the numbers are slightly different from our original paper (within 0.3% in mask AP).