RefineNet tensorflow 实现

• 使用ResNet作为特征提取器
• paper

文件部署

• ResNetV1： 残差网络
• RefineNet:
• Augmentation: 数据增强
• Data： 包含voc数据从图片转换为tfrecords形式的类和voc数据分装类
  • 注意先要对原始图片进行
• Train: 训练
• Test： 测试
• Predict： 预测

prepare

• download the pretrain model of resnet_v1_101.ckpt, you can download it from here
• download the pascal voc dataset
• some dependence like cv2, numpy and etc. recommend to install Anaconda

training

• first, run convert_pascal_voc_to_tfrecords.py to convert training data into .tfrecords, Or you can use the tfrecord I converted In BaiduYun.Currently, I only use the pascal voc 2012 for training.
• second, run python RefineNet/multi_gpu_train.py, also, you can change some hyper parameters in this file, like the batch size.

eval

• if you have already got a model, or just download the model I trained on pascal voc.model.
• put images in demo/ and run python RefineNet/demo.py

roadmap

• python2/3 compatibility
• Complete realization of refinenet model
• test on pascal voc, give the IoU result
• training on other datasets

• 参考github