/PSENet.pytorch

A pytorch re-implementation of PSENet: Shape Robust Text Detection with Progressive Scale Expansion Network

Primary LanguageC++GNU General Public License v3.0GPL-3.0

Shape Robust Text Detection with Progressive Scale Expansion Network

Requirements

  • pytorch 1.1
  • torchvision 0.3
  • pyclipper
  • opencv3
  • gcc 4.9+

Update

20190401

  1. add author loss, the results are compared in Performance

Download

resnet50 and resnet152 model on icdar 2015:

  1. bauduyun extract code: rxjf

  2. google drive

Data Preparation

follow icdar15 dataset format

img
│   1.jpg
│   2.jpg   
│		...
gt
│   gt_1.txt
│   gt_2.txt
|		...

Train

  1. config the trainroot,testrootin config.py
  2. use following script to run
python3 train.py

Test

eval.py is used to test model on test dataset

  1. config model_path, data_path, gt_path, save_path in eval.py
  2. use following script to test
python3 eval.py

Predict

predict.py is used to inference on single image

  1. config model_path, img_path, gt_path, save_path in predict.py
  2. use following script to predict
python3 predict.py

Performance

ICDAR 2015

only train on ICDAR2015 dataset with single NVIDIA 1080Ti

my implementation with my loss use adam and warm_up

Method Precision (%) Recall (%) F-measure (%) FPS(1080Ti)
PSENet-1s with resnet50 batch 8 81.13 77.03 79.03 1.76
PSENet-2s with resnet50 batch 8 81.36 77.13 79.18 3.55
PSENet-4s with resnet50 batch 8 81.00 76.55 78.71 4.43
PSENet-1s with resnet152 batch 4 85.45 80.06 82.67 1.48
PSENet-2s with resnet152 batch 4 85.42 80.11 82.68 2.56
PSENet-4s with resnet152 batch 4 83.93 79.00 81.39 2.99

my implementation with my loss use adam and MultiStepLR

Method Precision (%) Recall (%) F-measure (%) FPS(1080Ti)
PSENet-1s with resnet50 batch 8 83.39 79.29 81.29 1.76
PSENet-2s with resnet50 batch 8 83.22 79.05 81.08 3.55
PSENet-4s with resnet50 batch 8 82.57 78.23 80.34 4.43
PSENet-1s with resnet152 batch 4 85.33 79.87 82.51 1.48
PSENet-2s with resnet152 batch 4 85.36 79.73 82.45 2.56
PSENet-4s with resnet152 batch 4 83.95 78.86 81.33 2.99

my implementation with author loss use adam and warm_up

Method Precision (%) Recall (%) F-measure (%) FPS(1080Ti)
PSENet-1s with resnet50 batch 8 83.33 77.75 80.44 1.76
PSENet-2s with resnet50 batch 8 83.01 77.66 80.24 3.55
PSENet-4s with resnet50 batch 8 82.38 76.98 79.59 4.43
PSENet-1s with resnet152 batch 4 85.16 79.87 82.43 1.48
PSENet-2s with resnet152 batch 4 85.03 79.63 82.24 2.56
PSENet-4s with resnet152 batch 4 84.53S 79.20 81.77 2.99

my implementation with author loss use adam and MultiStepLR

Method Precision (%) Recall (%) F-measure (%) FPS(1080Ti)
PSENet-1s with resnet50 batch 8 83.93 79.48 81.65 1.76
PSENet-2s with resnet50 batch 8 84.17 79.63 81.84 3.55
PSENet-4s with resnet50 batch 8 83.50 78.71 81.04 4.43
PSENet-1s with resnet152 batch 4 85.16 79.58 82.28 1.48
PSENet-2s with resnet152 batch 4 85.13 79.15 82.03 2.56
PSENet-4s with resnet152 batch 4 84.40 78.71 81.46 2.99

official implementation use SGD and StepLR

Method Precision (%) Recall (%) F-measure (%) FPS(1080Ti)
PSENet-1s with resnet50 batch 8 84.15 80.26 82.16 1.76
PSENet-2s with resnet50 batch 8 83.61 79.82 81.67 3.72
PSENet-4s with resnet50 batch 8 81.90 78.23 80.03 4.51
PSENet-1s with resnet152 batch 4 82.87 78.76 80.77 1.53
PSENet-2s with resnet152 batch 4 82.33 78.33 80.28 2.61
PSENet-4s with resnet152 batch 4 81.19 77.13 79.11 3.00

examples

reference

  1. https://github.com/liuheng92/tensorflow_PSENet
  2. https://github.com/whai362/PSENet