BY Guangrun Wang
Base model: efficientnet-b1+aspp;
Big model: efficientnet-b3+aspp;
def forward(self, x):
size = (x.shape[2], x.shape[3])
x = self.efficientnet_model(x)
x = self.aspp(x)
x = self.conv_cls(x)
x = nn.Upsample(size, mode='bilinear', align_corners=True)(x)
return x
Machine: 8 titanX GPUs, each of which is 12G
Epochs: 50
Test-stage augmentation: [257, 385, 513, 641, 769]:
Test-stage augmentation: Left-right Flipping
Dataset: PASCAL VOC
Training set
1. TrainAug
2. 1/4 TrainAug
3. 1/4 TrainAug + 3/4 TrainAug (Auto annotated)
Test set:
1. VOC Val
Results:
1. TrainAug
Base model w/o multi-scale, w/o flipping:
IoU 0: 91.14
IoU 1: 70.37
IoU 2: 34.44
IoU 3: 82.24
IoU 4: 55.90
IoU 5: 66.08
IoU 6: 86.84
IoU 7: 77.51
IoU 8: 88.38
IoU 9: 32.21
IoU 10: 63.96
IoU 11: 47.69
IoU 12: 80.13
IoU 13: 76.14
IoU 14: 64.28
IoU 15: 77.74
IoU 16: 50.22
IoU 17: 75.06
IoU 18: 44.94
IoU 19: 69.88
IoU 20: 62.75
Mean IoU: 66.57
Base model w/ multi-scale, w/ flipping:
IoU 0: 92.75
IoU 1: 82.19
IoU 2: 39.18
IoU 3: 86.80
IoU 4: 65.64
IoU 5: 72.65
IoU 6: 89.32
IoU 7: 81.98
IoU 8: 91.11
IoU 9: 33.49
IoU 10: 72.78
IoU 11: 46.98
IoU 12: 84.85
IoU 13: 80.99
IoU 14: 71.17
IoU 15: 81.25
IoU 16: 56.47
IoU 17: 81.00
IoU 18: 44.05
IoU 19: 79.13
IoU 20: 65.85
Mean IoU: 71.41
2. 1/4 TrainAug
Base model w/o multi-scale, w/o flipping:
IoU 0: 88.29
IoU 1: 57.06
IoU 2: 28.15
IoU 3: 69.89
IoU 4: 53.46
IoU 5: 42.93
IoU 6: 77.99
IoU 7: 69.24
IoU 8: 79.15
IoU 9: 23.78
IoU 10: 30.61
IoU 11: 21.32
IoU 12: 60.54
IoU 13: 43.70
IoU 14: 58.71
IoU 15: 67.90
IoU 16: 21.41
IoU 17: 41.53
IoU 18: 31.56
IoU 19: 69.03
IoU 20: 56.48
Mean IoU: 52.03
Base model w/ multi-scale, w/ flipping:
IoU 0: 90.03
IoU 1: 65.20
IoU 2: 29.33
IoU 3: 74.74
IoU 4: 61.35
IoU 5: 47.67
IoU 6: 81.33
IoU 7: 73.16
IoU 8: 81.72
IoU 9: 23.15
IoU 10: 26.54
IoU 11: 18.05
IoU 12: 66.63
IoU 13: 44.61
IoU 14: 64.47
IoU 15: 69.18
IoU 16: 18.00
IoU 17: 48.87
IoU 18: 28.74
IoU 19: 73.61
IoU 20: 57.69
Mean IoU: 54.48
Big Model w/o multi-scale, w/o flipping:
IoU 0: 90.03
IoU 1: 67.75
IoU 2: 32.86
IoU 3: 72.25
IoU 4: 62.04
IoU 5: 67.90
IoU 6: 83.94
IoU 7: 76.87
IoU 8: 85.40
IoU 9: 27.14
IoU 10: 63.42
IoU 11: 34.75
IoU 12: 76.64
IoU 13: 62.85
IoU 14: 67.88
IoU 15: 76.35
IoU 16: 43.29
IoU 17: 63.98
IoU 18: 38.06
IoU 19: 76.52
IoU 20: 60.89
Mean IoU: 63.37
Big Model w/ multi-scale, w/ flipping:
IoU 0: 91.56
IoU 1: 76.36
IoU 2: 35.28
IoU 3: 77.22
IoU 4: 67.74
IoU 5: 72.06
IoU 6: 86.69
IoU 7: 80.83
IoU 8: 87.36
IoU 9: 28.88
IoU 10: 65.45
IoU 11: 36.12
IoU 12: 79.04
IoU 13: 65.09
IoU 14: 72.73
IoU 15: 79.33
IoU 16: 46.04
IoU 17: 67.98
IoU 18: 39.44
IoU 19: 80.38
IoU 20: 63.60
Mean IoU: 66.63
3. 1/4 TrainAug + 3/4 TrainAug (Auto annotated)
Base model w/ multi-scale, w/ flipping:
IoU 0: 91.13
IoU 1: 64.52
IoU 2: 36.17
IoU 3: 80.61
IoU 4: 68.63
IoU 5: 68.27
IoU 6: 87.86
IoU 7: 79.37
IoU 8: 87.80
IoU 9: 28.03
IoU 10: 68.95
IoU 11: 40.43
IoU 12: 80.06
IoU 13: 72.18
IoU 14: 74.71
IoU 15: 77.87
IoU 16: 46.76
IoU 17: 68.72
IoU 18: 44.14
IoU 19: 65.97
IoU 20: 62.45
Mean IoU: 66.41
Big model w/ multi-scale, w/ flipping:
IoU 0: 92.08
IoU 1: 68.89
IoU 2: 39.15
IoU 3: 82.89
IoU 4: 72.08
IoU 5: 73.92
IoU 6: 89.51
IoU 7: 79.23
IoU 8: 88.02
IoU 9: 27.94
IoU 10: 76.93
IoU 11: 42.52
IoU 12: 78.66
IoU 13: 73.64
IoU 14: 71.93
IoU 15: 81.44
IoU 16: 46.67
IoU 17: 70.39
IoU 18: 42.19
IoU 19: 78.96
IoU 20: 59.56
Mean IoU: 68.41