Branch Alignment Learning for Oriented Object Detection in Remote Sensing Images

$A$ is a set of all anchor boxes， $G_{G T}$ is a set of ground-truth boxes，is a set of classification predictions， $P_{r e g}$ is a set of regression predictions

Output：

$S_{p}$ is a set of positive samples， $S_{n}$ is a set of negative samples

1： for each $g \in G_{G T}$ do

2： $S_{c}$ ← compute IoU between g and $a \in A$ ， and select top Q anchors with the highest IoU as candidate samples $S_{c}$ ；

3： $P_{c l s, c}$ ， $P_{r e g, c}$ ← select the classification and regression predictions corresponding to the candidate sample $S_{c}$ ；

4： for each $p_{r e g}$ $\in P_{r e g, c}$ do

5： $I = I o U (p_{r e g}, g)$ ← compute IoU between $g$ and $p_{r e g}$ ；

6： end for

7： Create an empty set to save alignment degree： $A_{A D} \leftarrow \emptyset$ ；

8： for each $s_{c}$ $\in S_{c}$ do

9： $a d = c a l c u l a t e_A_{A D} (s_{c}, g)$ ← compute the $A_{A D}$ between $s_{c}$ and g；

10： $A_{A D} = A_{A D} ⋃ a_{a d}$ ；

11： end for

12： $S_{a l i} = {T_{t o p}}_{, v} (A_{A D})$ ← select top V candidate samples as alignment samples $S_{a l i}$ ；

13： $x_{g}, x_{a}$ ← transform $g$ and $S_{a l i}$ to the Gaussian distribution $x_{g}, x_{a}$ ；

14： $S_{p}$ ← compute the Gaussian similarity between $x_{g}$ and $x_{a}$ ， and define samples higher than the threshold as positive samples $S_{p}$ ；

15： end for

16： $S_{n} = A - S_{p}$ ；

17： return $S_{n}, S_{p}$ ；

Table 2. Compare mAP value of current mainstream one-stage, two-stage and anchor-free algorithms on DOTA-V1.0

View table

Table 2. Compare mAP value of current mainstream one-stage, two-stage and anchor-free algorithms on DOTA-V1.0

Method	PL	BD	BR	GTF	SV	LV	SH	TC	BC	ST	SBF	RA	HA	SP	HC	mAP
PSC^［17］	88.27	73.2	44.55	62.29	77.79	77.3	87.04	90.88	78.47	72.01	52.69	61.14	66.36	69.68	58.1	70.65
SCRDet^［18］	89.98	80.65	52.09	68.36	68.36	60.32	72.41	90.85	87.94	86.86	65.02	66.68	66.25	68.24	65.21	72.61
SASM^［19］	87.51	80.15	51.07	70.35	74.95	75.8	84.23	90.90	80.87	84.93	58.51	65.59	69.74	70.18	42.31	72.47
KFIoU^［20］	89.20	76.4	51.64	70.15	78.31	76.43	87.1	90.88	81.68	82.22	64.65	64.84	66.77	70.68	49.52	73.37
ARS-DETR^［21］	86.61	77.26	48.84	66.76	78.38	78.96	87.4	90.61	82.76	82.19	54.72	62.61	72.64	72.80	64.26	73.79
ReDet^［22］	88.94	78.07	51.19	72.76	74.26	78.08	87.44	90.84	80.79	78.59	60.85	64.22	76.84	72.79	54.85	74.03
RoI Trans^［23］	89.01	77.48	51.64	72.07	74.43	77.55	87.76	90.81	79.71	85.27	58.36	64.11	76.5	71.99	54.06	74.05
S2Anet^［3］	89.11	82.84	48.37	71.11	78.11	78.39	87.25	90.83	84.9	85.64	60.36	62.6	65.26	69.13	57.94	74.12
Baseline	88.67	77.62	41.81	58.17	74.58	71.64	79.11	90.29	82.13	74.32	54.75	60.6	62.57	69.67	60.64	68.43
Proposed	88.86	83.4	51.51	72.51	79.96	76.9	86.88	90.91	84.87	85.25	60.64	66.16	66.94	71.37	64.32	75.36

Table 3. Compare the mAP value of current mainstream algorithms on DOTA-V1.5

View table

Table 3. Compare the mAP value of current mainstream algorithms on DOTA-V1.5

Method	PL	BD	BR	GTF	SV	LV	SH	TC	BC	ST	SBF	RA	HA	SP	HC	CC	mAP
R3Det^［2］	71.99	74.61	48.4	61.49	45.89	64.96	76.59	90.82	71.6	68.7	48.35	66.47	68.31	65.95	60.15	16.27	62.54
DAL^［9］	80.80	80.99	46.43	64.25	49.52	68.02	70.92	90.73	74.28	72.49	48.44	67.6	60.49	65.77	49.86	24.04	64.19
Mask R-CNN	78.36	77.41	53.36	56.94	52.17	63.6	79.74	90.31	81.91	66.41	45.49	71.32	70.77	73.87	61.49	17.11	64.54
HTC	78.41	74.41	53.41	63.17	52.45	63.56	79.89	90.34	75.17	67.64	53.16	69.94	72.13	74.02	56.42	12.14	64.47
ReDet^［23］	79.51	82.63	53.81	69.82	52.76	75.64	87.82	90.83	75.81	68.78	49.11	71.65	75.57	75.17	58.29	15.36	67.66
Baseline	71.66	77.22	48.71	65.16	49.48	69.64	79.21	90.84	77.21	61.03	47.3	68.69	67.22	74.48	46.16	5.78	62.49
Proposed	79.97	82.74	53.23	67.82	53.16	76.72	86.58	90.92	77.82	70.71	47.29	69.92	71.59	74.68	62.75	26.36	68.26

Table 4. Accuracy comparison between this study and other algorithms at DIOR-R

View table

Table 4. Accuracy comparison between this study and other algorithms at DIOR-R

Method	Backbone	Image size /（pixel×pixel）	mAP /%
KFIoU^［20］	R-50	800×800	57.8
Gliding Vertex	R-50	800×800	60.1
GWD^［15］	R-101	800×800	60.3
RoI Trans^［22］	R-50	512×800	63.9
CGCDet	R-50	800×800	64.9
Baseline	R-50	800×800	57.6
Proposed	R-50	800×800	65.5

Table 5. Ablation study results of proposed each structure on DOTA-V1.0
View table
Table 5. Ablation study results of proposed each structure on DOTA-V1.0
Method Structure mAP /%
Baseline None 68.43
BFA-Net BAH 71.06
BFA-Net BAH+BAS 72.13
BFA-Net BAH+BAS+BASS 75.36

Table 6. Ablation study results of BAF layers N on DOTA-V1.0
View table
Table 6. Ablation study results of BAF layers N on DOTA-V1.0
N BAM+BASS Parameters GFLOPs mAP /%
Baseline ‒ 36.51 217.3 68.43
10 √ 38.71 209.24 74.72
8 √ 37.24 183.37 75.36
6 √ 35.76 157.7 74.83
4 √ 34.29 131.94 73.13

Table 7. Ablation study results of hyper-parameters in AD on DOTA-V1.0
View table
Table 7. Ablation study results of hyper-parameters in AD on DOTA-V1.0
$β$ $α$ mAP /% $β$ $α$ mAP /% $β$ $α$ mAP /% $β$ $α$ mAP /%
1/4 0.3 74.75 1/3 0.3 74.68 1/2 0.3 74.68 1 0.3 74.89
0.5 75.05 0.5 75.19 0.5 75.19 0.5 75.36
0.7 74.78 0.7 75.12 0.7 75.12 0.7 75.18
0.9 74.67 0.9 75.08 0.9 75.08 0.9 75.09

Table 8. Ablation study results of Gaussian similarity threshold on DOTA-V1.0
View table
Table 8. Ablation study results of Gaussian similarity threshold on DOTA-V1.0
$γ$ 0.3 0.4 0.5 0.6
mAP /% 74.51 75.36 75.02 74.79

Table 9. Portability analysis of BAM on DOTA-V1.0

View table

Table 9. Portability analysis of BAM on DOTA-V1.0

Method	BAM	Head/full parameters /Mbit	Head /full flops /Gbit	mAP /%
ATSS-O	‒	4.76/36.05	104.07/207.87	72.29
ATSS-O	√	4.83/37.29	79.67/183.47	74.13（ $↑$ 1.84）
FCOS-O	‒	4.77/31.92	104.92/206.91	72.76
FCOS-O	√	4.83/33.16	79.67/182.41	73.36（ $↑$ 0.60）
S2ANet	‒	5.47/38.61	119.57/197.62	74.12
S2ANet	√	4.97/38.51	98.72/179.58	74.75（ $↑$ 0.63）

Tools

Get Citation

Copy Citation Text

Hailong Zhang, Qiaolin Zeng, Jie Yang, Bowei Wang, Chengfang Wang. Branch Alignment Learning for Oriented Object Detection in Remote Sensing Images[J]. Laser & Optoelectronics Progress, 2025, 62(4): 0428005

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites