text
stringlengths
0
820
epoch 33. We use one of the SOTA OBB detection frameworks
— ORCN [126] to evaluate the performance of different pre-
trained backbones. We adopt the default hyper-parameters of
ORCN, which is implemented in OBBDetection2. Following
[126], the DOTA dataset is sampled and cropped to 1,024
×1,024 patches with a stride of 824, while the HRSC2016
images are scaled keeping the aspect ratio with the shorter side
equals to 800, and the length of the longer side is less than
or equal to 1333. Data augmentations during training include
random horizontal and vertical flipping. For convenience, the
original training and validation sets are merged for training,
2https://github.com/jbwang1997/OBBDetection
WANG et al. : EMPIRICAL STUDY OF REMOTE SENSING PRETRAINING 13
TABLE VIII
RESULTS OF THE ORCN DETECTION MODEL WITH DIFFERENT BACKBONES AND SOTA METHODS ON THE TESTING SET OF THE DOTA DATASET .†: THE
RESULT IS FROM AERIAL DETECTON [110]. ‡: THE RESULT IS FROM THE ORIGINAL ORCN PAPER .
Method Backbone mAPAP per category1
Ship ST BD TC BC GTF Bridge LV SV HC SP RA SBF Plane Harbor
One-stage
RetinaNet [111]† IMP-ResNet-50-FPN 68.43 79.11 74.32 77.62 90.29 82.18 58.17 41.81 71.64 74.58 60.64 69.67 60.60 54.75 88.67 62.57
DAL [112] IMP-ResNet-50-FPN 71.44 79.74 78.45 76.55 90.84 79.54 66.80 45.08 76.76 67.00 60.11 73.14 62.27 57.71 88.68 69.05
RSDet [113] IMP-ResNet-101-FPN 72.20 70.20 83.40 82.90 90.50 85.60 65.20 48.60 70.10 69.50 68.00 * 67.20 63.90 62.50 89.80 65.60
R3Det [114] IMP-ResNet-101-FPN 71.69 77.54 83.54 81.99 90.80 81.39 62.52 48.46 74.29 70.48 60.05 67.46 59.82 61.97 89.54 65.44
R3Det [114] IMP-ResNet-152-FPN 73.74 78.21 84.23 81.17 90.81 85.26 66.10 50.53 78.66 70.92 67.17 69.83 63.77 61.81 89.49 68.16
S2ANet [115] IMP-ResNet-50-FPN 74.12 87.25 85.64 82.84 90.83 84.90 71.11 48.37 78.39 78.11 57.94 69.13 62.60 60.36 89.11 65.26
S2ANet [115] IMP-ResNet-101-FPN 76.11 88.04 86.22 81.41 90.69 84.75 69.75 54.28 80.54 78.04 58.86 73.37 * 65.81 65.03 88.70 76.16
Two-stage
ICN [116] IMP-ResNet-101-FPN 68.16 69.98 78.20 74.30 90.76 79.06 70.32 47.70 67.82 64.89 50.23 64.17 62.90 53.64 81.36 67.02
Faster R-CNN [117] † IMP-ResNet-50-FPN 69.05 77.11 83.90 73.06 90.84 78.94 59.09 44.86 71.49 73.25 56.18 64.91 62.95 48.59 88.44 62.18
CAD-Net [118] IMP-ResNet-101-FPN 69.90 76.60 73.30 82.40 90.90 * 79.20 73.50 49.40 63.50 71.10 62.20 67.00 60.90 48.40 87.80 62.00
ROI Transformer [119] IMP-ResNet-101-FPN 69.56 83.59 81.46 78.52 90.74 77.27 75.92 43.44 73.68 68.81 47.67 58.93 53.54 58.39 88.64 62.83
SCRDet [120] IMP-ResNet-101-FPN 72.61 72.41 86.86 * 80.65 90.85 87.94 68.36 52.09 60.32 68.36 65.21 68.24 66.68 65.02 89.98 66.25
ROI Transformer†[119] IMP-ResNet-50-FPN 74.61 86.87 82.51 82.60 90.71 83.83 70.87 52.53 76.67 77.93 61.03 68.75 67.61 53.95 88.65 74.67
Gliding Vertex [121] IMP-ResNet-101-FPN 75.02 86.82 86.81 85.00 90.74 79.02 77.34 * 52.26 73.14 73.01 57.32 70.86 70.91 * 59.55 89.64 72.94
FAOD [122] IMP-ResNet-101-FPN 73.28 79.56 84.68 79.58 90.83 83.40 76.41 45.49 68.27 73.18 64.86 69.69 65.42 53.40 90.21 * 74.17
CenterMap-Net [123] IMP-ResNet-50-FPN 71.74 78.10 83.61 81.24 88.83 77.80 60.65 53.15 66.55 78.62 58.70 72.36 66.19 49.36 88.88 72.10
FR-Est [124] IMP-ResNet-101-FPN 74.20 86.44 83.56 81.17 90.82 84.13 70.19 50.44 77.98 73.52 60.55 66.72 66.59 60.64 89.63 70.59
Mask OBB [125] IMP-ResNet-50-FPN 74.86 85.57 85.05 85.09 * 90.37 82.08 72.90 51.85 73.23 75.28 66.33 69.87 68.39 55.73 89.61 71.61
ORCN‡[126] IMP-ResNet-50-FPN 75.87 88.20 * 84.68 82.12 90.90 * 87.50 70.86 54.78 83.00 78.93 52.28 68.84 67.69 63.97 89.46 74.94
ORCN‡[126] IMP-ResNet-101-FPN 76.28 87.52 85.33 83.48 90.90 * 85.56 76.92 55.27 82.10 74.27 57.28 70.15 66.82 65.51 * 88.86 74.36
ORCN IMP-ResNet-50-FPN 76.14 88.16 84.91 81.35 90.90* 87.43 71.35 54.86 83.03 79.04 58.14 69.05 66.67 63.39 89.58 74.19
ORCN SeCo-ResNet-50-FPN 70.02 86.33 81.31 73.32 90.88 79.46 67.07 49.94 76.48 76.15 49.71 65.32 58.55 41.31 88.64 65.90
ORCN RSP-ResNet-50-FPN 76.50 88.17 85.72 81.88 90.84 86.17 70.91 54.39 83.01 78.67 62.22 72.21 67.45 62.22 89.78 73.99
ORCN IMP-Swin-T-FPN 76.07 88.02 84.92 82.23 90.90* 87.42 74.37 52.25 83.55 77.99 63.07 69.30 65.99 57.70 89.48 73.88
ORCN RSP-Swin-T-FPN 76.12 87.83 84.84 79.74 90.86 85.90 74.50 52.91 84.02 78.96 57.36 70.61 67.33 62.90 89.54 74.45
ORCN IMP-ViTAEv2-S-FPN 77.38 88.14 86.35 83.50 90.90 * 87.51 75.38 53.42 85.15* 79.99* 66.03 66.12 70.91* 61.02 89.27 76.95*
ORCN RSP-ViTAEv2-S-FPN 77.72* 88.04 85.58 83.04 90.90 * 88.17* 75.16 55.85* 84.34 79.95 67.89 67.15 70.60 62.64 89.66 76.77
1ST: storage tank. BD: baseball diamond. TC: tennis court. BC: baseball court. GTF: ground track field. LV: large vehicle. SV: small vehicle. HC: helicopter.
SP: swimming pool. RA: roundabout. SBF: soccer ball field.
TABLE IX
RESULTS OF THE ORCN DETECTION MODEL WITH DIFFERENT
BACKBONES AND SOTA METHODS ON THE TESTING SET OF THE
HRSC2016 DATASET .†: THE RESULT IS FROM THE ORIGINAL ORCN
PAPER .
Method Backbone mAP
R2PN [127] IMP-VGG-16 79.6
RRD [128] IMP-VGG-16 84.3
FoRDet [129] IMP-VGG-16 89.9
R2CNN [130] IMP-ResNet-101 73.1
Rotated RPN [131] IMP-ResNet-101 79.1
ROI Transformer [119] IMP-ResNet-101-FPN 86.2
Gliding Vertex [121] IMP-ResNet-101-FPN 88.2
GRS-Det [132] IMP-ResNet-50-FPN 88.9
GRS-Det [132] IMP-ResNet-101-FPN 89.6
R3Det [114] IMP-ResNet-101-FPN 89.3
DAL [112] IMP-ResNet-101-FPN 89.8
AproNet [115] IMP-ResNet-101-FPN 90.0
S2ANet [115] IMP-ResNet-101-FPN 90.2
ORCN†[126] IMP-ResNet-50-FPN 90.4
CHPDet [133] Hourglass104 90.6*
ORCN IMP-ResNet-50-FPN 90.4
ORCN SeCo-ResNet-50-FPN 88.9
ORCN RSP-ResNet-50-FPN 90.3
ORCN IMP-Swin-T-FPN 89.7
ORCN RSP-Swin-T-FPN 90.0
ORCN IMP-ViTAEv2-S-FPN 90.4
ORCN RSP-ViTAEv2-S-FPN 90.4
while the original testing sets of DOTA and HRSC2016 are
separately used for evaluation. We report the mean average
precision (mAP) of all categories and the average precision
(AP) of each class on the corresponding testing set. All models
are trained on a single V100 GPU.
3) Experimental Results: Quantitative Results and Anal-
yses: Table VIII-IX show the results of OBB detection
experiments. On the challenging DOTA dataset, it can be
seen that using the advanced ORCN framework, the models
whose backbone is either ResNet-50 or Swin-T performs well,
although the mAPs of Swin-T models are slightly lower than
the ResNet models. The ViTAEv2-S, which is a kind of vision
transformer network that is introduced the inductive biasesincluding the locality and scale-invariance characteristics of
CNN, obtains amazing performance that improves the ORCN
baseline by nearly 2% mAP. Another point needed to be
noticed is the performance of RSP weights on these three
backbones all outperforms their ImageNet pretrained coun-
terparts. These results support our previous claims that the
granularity of the representation required for the detection
task is closer to that for the scene recognition task compared
with the segmentation task. Thus, the performance difference
between RSP and IMP in the detection experiments aligns with
the results in the scene recognition experiments.