Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
AIST: Satellite Imagery Object Detection
Yk
Aizawa Laboratory
The University of Tokyo
Self-introduction
• Name: 郁 (Yu) 青 (Qing)
• Hometown: Shanghai, China
• Affiliation: Aizawa Laboratory, The University of Tokyo
• Grade: M1
2
AIST: Satellite Imagery Object Detection
• Binary classification: golf course or not
3
AIST: Satellite Imagery Object Detection
• Binary classification: golf course or not
4
Golf Course
Not Golf Course
AIST: Satellite Imagery Object Detection
• Binary classification: golf course or not
5
Golf Course
Not Golf Course
Data
• Train: 296,182 images (Positive: 12,442 Negative: 283,740)
• Test: 428,697 images
• Size: 7 channel x 32 pix x 32 pix
• Model evaluation:
6
𝐼𝑜𝑈 =𝑇𝑃
𝐹𝑁 + 𝑇𝑃 + 𝐹𝑃
Method
7
7 channel image
DCNN
…. Score
• Pon de deep
>threshold: Positive<threshold: Negative
Deep Convolution Neural Network
• DCNN: PreResNet-110 pretrained on CIFAR-100
8
original pre-activation
Residual Unit
Deep Convolution Neural Network
• DCNN: PreResNet-110 pretrained on CIFAR-100
• RGB Image: 3 channel
• Satellite Image: 7 channel
➢Modified first convolution layer
9
Deep Convolution Neural Network
• DCNN: PreResNet-110 pretrained on CIFAR-100
• Modified first convolution layer
10
3 x 16 x 3 x 3
in channel
out channel
Filter size
height
width
Normal 3 channel image
Deep Convolution Neural Network
• DCNN: PreResNet-110 pretrained on CIFAR-100
• Modified first convolution layer
11
3 x 16 x 3 x 3
in channel
Filter size 1 x 16 x 3 x 3mean filter
Normal 3 channel image
Deep Convolution Neural Network
• DCNN: PreResNet-110 pretrained on CIFAR-100
• Modified first convolution layer
12
3 x 16 x 3 x 3
in channel
Filter size 1 x 16 x 3 x 3mean filter
7 x 16 x 3 x 3
in channel
copy
7 channel satellite imageNormal 3 channel image
Data Augmentation
13
• Train: RandomTranslateWithReflect, RandomFlip,RandomErasing,Mixup
• Test: None
Hyperparameter
14
• Optimizer: SGD with Nesterov Momentum
• Learning rate: cosine annealing from 0.01
• Momentum: 0.9
• Weight decay: 1e-3
• Epoch: 100
• Batch size: 400
• Threshold: 0.7
• Training needs 12 hours on 1 Titan Xp
Rank UsernamePublicScore
PrivateScore
1 Kenmatsu4 0.83743 0.81203
2 Yk 0.83142 0.82147
Ranked in Top 2 by this model
Final Model
15
• Ensemble
Average Score
0.86045
0.83142
Result
16
Rank UsernamePublicScore
PrivateScore
Submissions Submitted
1 Yk 0.86045 0.83462 482018-04-26
20:07:07
2 Kenmatsu4 0.83743 0.81203 952018-04-26
20:40:52
3 bacon 0.80527 0.80711 22018-04-23
14:37:52
4 4Ui_iUrz1 0.81092 0.79831 842018-04-26
20:05:16
5 coz.a 0.78761 0.79174 402018-04-24
22:58:04
Summary
• Power is justice!!!
17