refinenet

a tensorflow implement of refinenet. RefineNet: Multi-Path Refinement Networks for High-Resolution Semantic Segmentation

Introduction

this is a tensorflow implement of refinenet discribed in arxiv:1611.06612.I have not finished it yet, this is just a demo, but the model is already able to work.

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

some result