/Gaofen-challenge

The top-5 winning solution (5/152) in 2021 Gaofen Challenge.

Primary LanguagePythonGNU General Public License v3.0GPL-3.0

2021 Gaofen Challenge

The fifth winning solution (5/152) in the Object Recognition in Sar Images, 2021 Gaofen Challenge on Automated High-Resolution Earth Observation Image Interpretation.

image-20230327142923716

Members

Caihua kong, Longgang Dai, Zhentao Fan, Xiang Chen.

Solution

  • 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

Experiment

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

How to use

  1. 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).

  2. Prepare your environment

    Our code is implemented based on mmyolo, see mmyolo for environment installation.

  3. Train and Test

    Please see command.md.

  4. Test for submit

    Run python run.py /input_path /out_path.

  5. Docker submit

    Please see my_blog

Detections

image-20230327152202202