The fifth winning solution (5/152) in the Object Recognition in Sar Images, 2021 Gaofen Challenge on Automated High-Resolution Earth Observation Image Interpretation.
Caihua kong, Longgang Dai, Zhentao Fan, Xiang Chen.
-
On-line date augmentation
We use random combination of affine transformation, flip, scaling, optical distortion for data augmentation. -
Copy-Paste off-line data augmention
-
Multi-scale training and testing
The training images are resized into sizes of 600, 800, and 1024 for training and testing. -
WBF
-
PPYOLOE
版本号 | val@mAP | test@mAP | val@time(s)/TTA | test@time(s) |
---|---|---|---|---|
V1.0.13 | - | 70.3977 | - | 764 |
v2.2.2 | 71.30 | 67.3279 | 67.32 | 175 |
v2.2.4 | 76.31 | 67.9190 | 82 | 125 |
v2.2.5 | 78.60 | - | 33 | None |
v2.2.6 | - | 67.5479 | 80 | 121(x1.5) |
v2.2.7 | 80.7 | 68.4315 | 40(68s) | 127 |
v2.2.8 | 86 | 69.7269 | 31(74s) | 146(x1.9) |
v2.2.9 | 83.7 | 67.6951 | 29(83) | 57 |
For detailed records, please refer to the experiment.md
-
Prepare sar dataset
Please execute the script under
dataset_converter
to convert the data format from xml format to coco format.Note: the official only gave the training set, and the verification set is mainly defined by itself (8:2).
-
Prepare your environment
Our code is implemented based on mmyolo, see mmyolo for environment installation.
-
Train and Test
Please see
command.md
. -
Test for submit
Run
python run.py /input_path /out_path
. -
Docker submit
Please see my_blog