Cascade R-CNN: Delving into High Quality Object Detection

by Zhaowei Cai and Nuno Vasconcelos

This repository is written by Zhaowei Cai at UC San Diego, on the base of Detectron @ e8942c8.

Introduction

This repository re-implements Cascade R-CNN on the base of Detectron. Very consistent gains are available for all tested models, regardless of baseline strength.

Please follow Detectron on how to install and use Detectron-Cascade-RCNN.

It is also recommended to use our original implementation, cascade-rcnn based on Caffe, and the third-party implementation, mmdetection based on PyTorch and tensorpack based on TensorFlow.

Citation

If you use our code/model/data, please cite our paper:

@inproceedings{cai18cascadercnn,
  author = {Zhaowei Cai and Nuno Vasconcelos},
  Title = {Cascade R-CNN: Delving into High Quality Object Detection},
  booktitle = {CVPR},
  Year  = {2018}
}

and Detectron:

@misc{Detectron2018,
  author =       {Ross Girshick and Ilija Radosavovic and Georgia Gkioxari and
                  Piotr Doll\'{a}r and Kaiming He},
  title =        {Detectron},
  howpublished = {\url{https://github.com/facebookresearch/detectron}},
  year =         {2018}
}

Benchmarking

End-to-End Faster & Mask R-CNN Baselines

backbone	type	lr schd	im/ gpu	box AP	box AP50	box AP75	mask AP	mask AP50	mask AP75	download links
R-50-FPN-baseline	Faster	1x	2	36.7	58.4	39.6	-	-	-	model | boxes
R-50-FPN-cascade	Faster	1x	2	40.9	59.0	44.6	-	-	-	model | boxes
R-101-FPN-baseline	Faster	1x	2	39.4	61.2	43.4	-	-	-	model | boxes
R-101-FPN-cascade	Faster	1x	2	42.8	61.4	46.1	-	-	-	model | boxes
X-101-64x4d-FPN-baseline	Faster	1x	1	41.5	63.8	44.9	-	-	-	model | boxes
X-101-64x4d-FPN-cascade	Faster	1x	1	45.4	64.0	49.8	-	-	-	model | boxes
X-101-32x8d-FPN-baseline	Faster	1x	1	41.3	63.7	44.7	-	-	-	model | boxes
X-101-32x8d-FPN-cascade	Faster	1x	1	44.7	63.7	48.8	-	-	-	model | boxes
R-50-FPN-baseline	Mask	1x	2	37.7	59.2	40.9	33.9	55.8	35.8	model | boxes | masks
R-50-FPN-cascade	Mask	1x	2	41.3	59.6	44.9	35.4	56.2	37.8	model | boxes | masks
R-101-FPN-baseline	Mask	1x	2	40.0	61.8	43.7	35.9	58.3	38.0	model | boxes | masks
R-101-FPN-cascade	Mask	1x	2	43.3	61.7	47.2	37.1	58.6	39.8	model | boxes | masks
X-101-64x4d-FPN-baseline	Mask	1x	1	42.4	64.3	46.4	37.5	60.6	39.9	model | boxes | masks
X-101-64x4d-FPN-cascade	Mask	1x	1	45.9	64.4	50.2	38.8	61.3	41.6	model | boxes | masks
X-101-32x8d-FPN-baseline	Mask	1x	1	42.1	64.1	45.9	37.3	60.3	39.5	model | boxes | masks
X-101-32x8d-FPN-cascade	Mask	1x	1	45.8	64.1	50.3	38.6	60.6	41.5	model | boxes | masks

Mask R-CNN with Bells & Whistles

backbone	type	lr schd	im/ gpu	box AP	box AP50	box AP75	mask AP	mask AP50	mask AP75	download links
X-152-32x8d-FPN-IN5k-baseline	Mask	s1x	1	48.1	68.3	52.9	41.5	65.1	44.7	model | boxes | masks
[above without test-time aug.]				45.2	66.9	49.7	39.7	63.5	42.4
X-152-32x8d-FPN-IN5k-cascade	Mask	s1x	1	50.2	68.2	55.0	42.3	65.4	45.8	model | boxes | masks
[above without test-time aug.]				48.1	66.7	52.6	40.7	63.7	43.8

Faster & Mask R-CNN with GN

backbone	type	lr schd	im/ gpu	box AP	box AP50	box AP75	mask AP	mask AP50	mask AP75	download links
R-50-FPN-GN-baseline	Faster	1x	2	38.4	59.9	41.7	-	-	-	model | boxes
R-50-FPN-GN-cascade	Faster	1x	2	42.2	60.6	45.8	-	-	-	model | boxes
R-101-FPN-GN-baseline	Faster	1x	2	39.9	61.3	43.3	-	-	-	model | boxes
R-101-FPN-GN-cascade	Faster	1x	1	43.8	62.2	47.6	-	-	-	model | boxes
R-50-FPN-GN-baseline	Mask	1x	2	39.2	60.5	42.9	34.9	57.1	36.9	model | boxes
R-50-FPN-GN-cascade	Mask	1x	1	42.9	60.7	46.6	36.6	57.7	39.2	model | boxes | masks
R-101-FPN-GN-baseline	Mask	1x	2	41.1	62.1	45.1	36.3	58.9	38.5	model | boxes | masks
R-101-FPN-GN-cascade	Mask	1x	1	44.8	62.8	48.8	38.0	59.8	40.8	model | boxes | masks