update readme

argman · Aug 5, 2017 · b33f2da · b33f2da
1 parent 42295df
commit b33f2da
Showing 1 changed file with 22 additions and 4 deletions.
diff --git a/readme.md b/readme.md
@@ -1,8 +1,26 @@
-# EAST : An Efficient and Accurate Scene Text Detector
+# EAST: An Efficient and Accurate Scene Text Detector
 
 ### Introduction
-This is a tensorflow implemention of EAST. I only reimplement the RBOX part of the paper, which achieves an F1 score
-of 80.8 on the ICDAR 2015 dataset (which is about two points better than the result of pvanet in the paper, see http://rrc.cvc.uab.es/?ch=4&com=evaluation&task=1). The running speed is about 150ms (network) + 300ms (NMS) per image on a K40 card. The nms part is too slow because of the use of shapely in python, and can be further improved.
+This is a tensorflow re-implementation of [EAST: An Efficient and Accurate Scene Text Detector](https://arxiv.org/abs/1704.03155v2).
+The features are summarized blow:
++ Only **RBOX** part is implemented.
++ A fast Locality-Aware NMS in C++ provided by the paper's author.
++ The pre-trained model provided achieves **80.83** F1-score on ICDAR 2015
+	Incidental Scene Text Detection using only images from ICDAR 2015 and 2013.
+  see [here](http://rrc.cvc.uab.es/?ch=4&com=evaluation&view=method_samples&task=1&m=29855&gtv=1) for the detailed results.
++ Differences from original paper
+	+ Use ResNet-50 rather than PVANET
+	+ Use dice loss (optimize IoU of segmentation) rather than balanced cross entropy
+	+ Use linear learning rate decay rather than staged learning rate decay
++ Speed
+	+ Now
+		+ Graphic card: GTX 1080 Ti
+		+ Network fprop: ~50 ms
+		+ NMS (C++): ~6ms
+	+ Then
+		+ Graphic card: K40
+		+ Network fprop: ~150 ms
+		+ NMS (python): ~300ms
 
 Thanks for the author's ([@zxytim](https://github.com/zxytim)) help!
 Please site his [paper](https://arxiv.org/abs/1704.03155v2) if you find this useful.
@@ -15,7 +33,7 @@ Please site his [paper](https://arxiv.org/abs/1704.03155v2) if you find this use
 5. [Examples](#examples)
 
 ### Installation
-1. I think any version of tensorflow version > 1.0 should be ok.
+1. Any version of tensorflow version > 1.0 should be ok.
 
 ### Download
 1. Models trained on ICDAR 2013 (training set) + ICDAR 2015 (training set): [BaiduYun link](http://pan.baidu.com/s/1jHWDrYQ)