Faster R-CNN画PR曲线

参考：https://github.com/rbgirshick/py-faster-rcnn/issues/670

原地址https://blog.csdn.net/hongxingabc/article/details/80064574

在pascal_voc.py里添加几行代码即可：

1，文件头部：

 

1. import matplotlib.pyplot as plt

2. import pylab as pl

3. from sklearn.metrics import precision_recall_curve

4. from itertools import cycle

2，_do_python_eval函数：
def _do_python_eval(self, output_dir='output'):

1. annopath = os.path.join(

2. self._devkit_path,

3. 'VOC' + self._year,

4. 'Annotations',

5. '{:s}.xml')

6. imagesetfile = os.path.join(

7. self._devkit_path,

8. 'VOC' + self._year,

9. 'ImageSets',

10. 'Main',

11. self._image_set + '.txt')

12. cachedir = os.path.join(self._devkit_path, 'annotations_cache')

13. aps = []

14. # The PASCAL VOC metric changed in 2010

15. use_07_metric = True if int(self._year) < 2010 else False

16. print('VOC07 metric? ' + ('Yes' if use_07_metric else 'No'))

17. if not os.path.isdir(output_dir):

18. os.mkdir(output_dir)

19. for i, cls in enumerate(self._classes):

20. if cls == '__background__':

21. continue

22. filename = self._get_voc_results_file_template().format(cls)

23. rec, prec, ap = voc_eval(

24. filename, annopath, imagesetfile, cls, cachedir, ovthresh=0.5,

25. use_07_metric=use_07_metric)

26. aps += [ap]

27. pl.plot(rec, prec, lw=2,

28. label='Precision-recall curve of class {} (area = {:.4f})'

29. ''.format(cls, ap))

30. print(('AP for {} = {:.4f}'.format(cls, ap)))

31. with open(os.path.join(output_dir, cls + '_pr.pkl'), 'wb') as f:

32. pickle.dump({'rec': rec, 'prec': prec, 'ap': ap}, f)

33.  

34. pl.xlabel('Recall')

35. pl.ylabel('Precision')

36. plt.grid(True)

37. pl.ylim([0.0, 1.05])

38. pl.xlim([0.0, 1.0])

39. pl.title('Precision-Recall')

40. pl.legend(loc="upper right")

41. plt.show()

42.  

43. print(('Mean AP = {:.4f}'.format(np.mean(aps))))

44. print('~~~~~~~~')

45. print('Results:')

46. for ap in aps:

47. print(('{:.3f}'.format(ap)))

48. print(('{:.3f}'.format(np.mean(aps))))

49. print('~~~~~~~~')

50. print('')

51. print('--------------------------------------------------------------')

52. print('Results computed with the **unofficial** Python eval code.')

53. print('Results should be very close to the official MATLAB eval code.')

54. print('Recompute with `./tools/reval.py --matlab ...` for your paper.')

55. print('-- Thanks, The Management')

56. print('--------------------------------------------------------------')

然后运行test_net.py，就可以得到如下图的PR曲线。如果想比较多条曲线，可以先把rec, prec数据存起来再画图。