Object Detection Made Simpler by Eliminating Heuristic NMS

PyTorch Implementation for Our Paper: "Object Detection Made Simpler by Eliminating Heuristic NMS"

Requirements

• Python 3.7
• PyTorch 1.5.1
• mmdetectoin

Usage:

The code is being submitted to the company for open source review.

PSS for nms-free Object Detection:

End-to-End Training

Model	Backbone	lr sched	mAP (COCO2017 val)	link
FCOS	R50	3x	42.0	model
ATSS	R50	3x	42.8	model
FCOSpss	R50	3x	42.3
ATSSpss	R50	3x	42.6
VFNETpss	R50	3x	44.0
FCOSpss	R101	3x	44.1
ATSSpss	R101	3x	44.2
VFNETpss	R101	3x	45.7
FCOSpss	X-101-DCN	2x	47.0
ATSSpss	X-101-DCN	2x	47.3
VFNETpss	X-101-DCN	2x	49.3
FCOSpss	R2N-101-DCN	2x	48.2
ATSSpss	R2N-101-DCN	2x	48.6
VFNETpss	R2N-101-DCN	2x	50.0

NOTE: All models are trained with multi-scale training schedule of ‘[480, 800] $\times$ 1333’.

Two-Step Training

If we have a pretrained model, only finetuning the PSS head can save the training time.

Model	Backbone	MS Training	lr sched	mAP pretrain model (w NMS)	mAP finetuned PSS (w/o NMS)
GFocalV2pss	R50	Yes	12	43.9	43.3
GFocalV2pss	X-101-32x4d-DCN	Yes	12	48.8	48.2
GFocalV2pss	R2N-101-DCN	Yes	12	49.9	49.2

PSS for nms-free Instance Segmentation

End-to-End Training

Model	Backbone	lr sched	bbox mAP	segm mAP	link
CondInst	R50	3x	41.9	37.5
CondInstpss	R50	3x	41.2	36.7
CondInst + sem	R50	3x	42.6	38.2
CondInstpss + sem	R50	3x	42.3	37.7
CondInst	R101	3x	43.3	38.6
CondInstpss	R101	3x	43.1	38.2
CondInst + sem	R101	3x	44.6	39.8
CondInstpss + sem	R101	3x	44.1	39.3

NOTE: All models are trained with multi-scale training schedule of ‘[640, 800] $\times$ 1333’.

Citation

If you use the package in your research, please cite our paper:

@misc{zhou2021object,
      title={Object Detection Made Simpler by Eliminating Heuristic NMS}, 
      author={Qiang Zhou and Chaohui Yu and Chunhua Shen and Zhibin Wang and Hao Li},
      year={2021},
      eprint={2101.11782},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

PS

团队长期招聘中，社招/实习/校招，我们都要 ! 欢迎投递简历，邮箱: zhouqiang@zju.edu.cn