Aerial-Semantic-Segmentation-Drone

dataset: https://www.kaggle.com/bulentsiyah/semantic-drone-dataset

Multi-GPU:

       Example:
       config:{'strategy': 
       tf.distribute.MirroredStrategy(devices=[
                                               "/gpu:0", 
                                               "/gpu:1", 
                                               "/gpu:2"])
       CUDA_VISIBLE_DEVICES="2,3,4" python train_v3.py

Single-GPU:

       Example:
       CUDA_VISIBLE_DEVICES="2" python train.py 

TODO：

       deeplabv3的tf.estimator下多GPU训练（改为MirroredStrategy）√
       复现的deeplabv3+在料口对齐上的效果 mIOU=0.9983 √
       MASK-RCNN源码（instance-segmentation）√
       增加图片强化：缩放、crop等
       预测、评估的代码