train-aeroscapes-with-gluon

无人机视角语义分割数据集 aeroscapes 的训练及预测 | train with drone view dataset aeroscapes and predict

quick glance: https://www.youtube.com/watch?v=yYhGy2Akg2o&list=PLM1v95K5B1ntVsYvNJIxgRPppngrO_X4s

环境 | depends on

• mxnet,gluon-cv

训练 | training

1. 准备数据集 | prepare dataset

数据集的谷歌网盘链接在这里：https://github.com/ishann/aeroscapes | Google Drive link of dataset here

解压并将数据集放到文件夹 ./aeroscapes | unzip and put aeroscapes dataset inside folder ./aeroscapes

train-aeroscapes-with-gluon
    aeroscapes/
        JPEGImages/
            3269 RGB images.
        SegmentationClass/
            3269 ground-truth segmentation masks.
        Visualizations/
            3269 RGB ground-truth segmentation visualizations.
        ImageSets/
            Training and validation splits for data.

然后放入数语义分析据集的txt文件 | then prepare Segmentation dataset txt

mv Segmentation/ aeroscapes/ImageSets/

2. 运行训练 | run training

python3 train_fcn.py

参数调整参考 train_fcn.py | params refer train_fcn.py

生成模型参数文件路径： runs/pascal_aug/fcn/default/checkpoint.params | output weights here

pretrained params(google drive): https://drive.google.com/file/d/1__mNQFFjCMonBI9tktN80Erx-e6blCSa/view?usp=sharing

检测 | predict

cp runs/pascal_aug/fcn/default/checkpoint.params ./
python3 predict.py

输入 | input：

输出 | output：