DOTA-DOAI

Abstract

This repo is the codebase for our team to participate in DOTA related competitions, including rotation and horizontal detection. We mainly use FPN-based two-stage detector, and it is completed by YangXue and YangJirui.

Performance

DOTA1.0 (Task1)

Model	Backbone	Training data	Val data	mAP	Model Link	Tricks	lr schd	Data Augmentation	GPU	Image/GPU	Configs
FPN (baseline)	ResNet50_v1 (600,800,1024)->800	DOTA1.0 trainval	DOTA1.0 test	69.35	model	No	1x	No	2X GeForce RTX 2080 Ti	1	cfgs_dota1.0_res50_v2.py
FPN	ResNet50_v1d (600,800,1024)->800	DOTA1.0 trainval	DOTA1.0 test	70.87	model	+InLD	1x	No	2X GeForce RTX 2080 Ti	1	cfgs_dota1.0_res50_v3.py
FPN	ResNet152_v1d (600,800,1024)->MS	DOTA1.0 trainval	DOTA1.0 test	76.20 (76.54)	model	ALL	2x	Yes	2X GeForce RTX 2080 Ti	1	cfgs_dota1.0_res152_v1.py

DOTA1.0 (Task2)

Model	Backbone	Training data	Val data	mAP	Model Link	Tricks	lr schd	Data Augmentation	GPU	Image/GPU	Configs
FPN (baseline)	ResNet50_v1 (600,800,1024)->800	DOTA1.0 trainval	DOTA1.0 test	76.03	model	No	1x	No	2X Quadro RTX 8000	1	cfgs_dota1.0_res50_v2.py
FPN (memory consumption)	ResNet152_v1d (600,800,1024)->MS	DOTA1.0 trainval	DOTA1.0 test	81.23	model	ALL	2x	Yes	2X Quadro RTX 8000	1	cfgs_dota1.0_res152_v1.py

Visualization

Performance of published papers on DOTA datasets

DOTA1.0 (Task1)

Model	Backbone	mAP	Paper Link	Code Link	Remark	Recommend
FR-O (DOTA)	ResNet101	52.93	CVPR2018	MXNet	DOTA dataset, baseline	✅
IENet	ResNet101	57.14	arXiv:1912.00969	-	anchor free
R2CNN	ResNet101	60.67	arXiv:1706.09579	TF	scene text, multi-task, different pooled sizes, baseline	✅
RRPN	ResNet101	61.01	TMM arXiv:1703.01086	TF	scene text, rotation proposals, baseline	✅
RetinaNet-H	ResNet101	64.73	arXiv:1908.05612	TF	single stage, baseline	✅
Axis Learning	ResNet101	65.98	Remote Sensing	-	single stage, anchor free	✅
ICN	ResNet101	68.16	ACCV2018	-	image cascade, multi-scale	✅
RADet	ResNeXt101	69.09	Remote Sensing	-	enhanced FPN, mask rcnn
RoI Transformer	ResNet101	69.56	CVPR2019	MXNet, Pytorch	roi transformer	✅
P-RSDet	ResNet101	69.82	arXiv:2001.02988	-	anchor free, polar coordinates	✅
CAD-Net	ResNet101	69.90	TGRS arXiv:1903.00857	-	attention
O2-DNet	Hourglass104	71.04	arXiv:1912.10694	-	anchor free	✅
SCRDet	ResNet101	72.61	ICCV2019	TF:R2CNN++, IoU-Smooth L1	attention, angular boundary problem	✅
SARD	ResNet101	72.95	Access	-	IoU-based weighted loss
FADet	ResNet101	73.28	ICIP2019	-	attention
MFIAR-Net	ResNet152	73.49	Sensors	-	feature attention, enhanced FPN
R3Det	ResNet152	73.74	arXiv:1908.05612	TF	refined single stage, feature alignment	✅
RSDet	ResNet152	74.10	arXiv:1911.08299	-	quadrilateral bbox, angular boundary problem	✅
Gliding Vertex	ResNet101	75.02	TPAMI arXiv:1911.09358	Pytorch	quadrilateral bbox	✅
Mask OBB	ResNeXt-101	75.33	Remote Sensing	-	attention, multi-task	✅
FFA	ResNet101	75.7	ISPRS	-	enhanced FPN, rotation proposals
APE	ResNeXt-101(32x4)	75.75	arXiv:1906.09447	-	length independent IoU (LIIoU)	✅
CSL	ResNet152	76.17 / 70.29	arXiv:2003.05597	TF:CSL_RetinaNet	angular boundary problem	✅
OWSR	Ensemble (ResNet101 + ResNeXt101 + mdcn-ResNet101)	76.36	CVPR2019 WorkShop TGRS	-	enhanced FPN
R3Det++	ResNet152	76.56	arXiv:2004.13316	TF	refined single stage, feature alignment, denoising	✅
SCRDet++	ResNet101	76.81	arXiv:2004.13316	TF	angular boundary problem, denoising	✅

DOTA1.0 (Task2)

Model	Backbone	mAP	Paper Link	Code Link	Remark	Recommend
FR-H (DOTA)	ResNet101	60.46	CVPR2018	MXNet	DOTA dataset, baseline	✅
SBL	ResNet50	64.77	arXiv:1810.08103	-	single stage
FMSSD	VGG16	72.43	TGRS	-	IoU-based weighted loss, enhanced FPN
ICN	ResNet101	72.45	ACCV2018	-	image cascade, multi-scale	✅
IoU-Adaptive R-CNN	ResNet101	72.72	Remote Sensing	-	IoU-based weighted loss, cascade
EFR	VGG16	73.49	Remote Sensing	Pytorch	enhanced FPN
SCRDet	ResNet101	75.35	ICCV2019	TF	attention, angular boundary problem	✅
FADet	ResNet101	75.38	ICIP2019	-	attention
MFIAR-Net	ResNet152	76.07	Sensors	-	feature attention, enhanced FPN
Mask OBB	ResNeXt-101	76.98	Remote Sensing	-	attention, multi-task	✅
A2RMNet	ResNet101	78.45	Remote Sensing	-	attention, enhanced FPN, different pooled sizes
OWSR	Ensemble (ResNet101 + ResNeXt101 + mdcn-ResNet101)	78.79	CVPR2019 WorkShop TGRS	-	enhanced FPN
DM-FPN	ResNet-Based	79.27	Remote Sensing	-	enhanced FPN
SCRDet++	ResNet101	79.35	arXiv:2004.13316	TF	denoising	✅

DOTA1.5 (Task1)

Model	Backbone	mAP	Paper Link	Code Link	Remark	Recommend
APE	ResNeXt-101(32x4)	78.34	arXiv:1906.09447	-	length independent IoU (LIIoU)	✅
OWSR	Ensemble (ResNet101 + ResNeXt101 + mdcn-ResNet101)	76.60	CVPR2019 WorkShop	-	enhanced FPN

DOTA1.5 (Task2)

Model	Backbone	mAP	Paper Link	Code Link	Remark	Recommend
OWSR	Ensemble (ResNet101 + ResNeXt101 + mdcn-ResNet101)	79.50	CVPR2019 WorkShop	-	enhanced FPN

Related Articles

Model	Paper Link	Code Link	Remark	Recommend
SSSDET	ICIP2019 arXiv:1909.00292	-	vehicle detection, lightweight
AVDNet	GRSL arXiv:1907.07477	-	vehicle detection, small object
ClusDet	ICCV2019	Caffe2	object cluster regions	✅
DMNet	CVPR2020 WorkShop	-	object cluster regions	✅
OIS	arXiv:1911.07732	related Pytorch code	Oriented Instance Segmentation	✅

Dataset

Some remote sensing related object detection dataset statistics are in DATASET.md