Parallel SIFT features extraction

[Thomas]

Dear List,

I get wrong features when running the jython script below (inspired from here, line 470) to extract SIFT features in parallel from several images present on disk. I can see that the features are wrong as with more than 2 threads, some images give erroneously large amounts of features (see logs below, images 37,39 for example).

Does somebody know whether I am doing something wrong ?

Thanks for your help

Thomas

Windows Server 2012, Fiji 64bit

import ij

from ij import IJ

from mpicbg.ij import SIFT

from mpicbg.imagefeatures import FloatArray2DSIFT

from java.util import HashSet

from java.lang import Thread

from java.util.concurrent.atomic import AtomicInteger

from java.lang import Runtime

paths = xxx # list containing the paths of the images

nImages = len(paths)

atomicI = AtomicInteger(0)

threads = []

def getFeatures():

floatSIFT = FloatArray2DSIFT(param)

ijSIFTLowRes = SIFT(floatSIFT)

while atomicI.get() < (nImages - 1):

k = atomicI.getAndIncrement()

if k < nImages:

imp = IJ.openImage(paths[k])

ip = imp.getProcessor()

features = HashSet()

ijSIFTLowRes.extractFeatures(ip, features)

IJ.log( str(k) + ': ' + str(features.size()) + ' features extracted' )

# if features.size() > 3000:

# IJ.log(str(k) + ': ' + '####################################SOMETHING WRONG##################')

imp.close()

del features

# Creating and starting the threads

for p in range(int (Runtime.getRuntime().availableProcessors() * 0.9 )):

# for p in range(1):

thread = Thread(getFeatures)

threads.append(thread)

thread.start()

IJ.log('Thread ' + str(p) + ' started')

# Joining the threads

for thread in threads:

thread.join()

IJ.log('Thread joined')

IJ.log('Computing features done')

Log 1 Thread 

Starting ...

Thread 0 started

0: 924 features extracted

1: 980 features extracted

2: 988 features extracted

3: 873 features extracted

4: 985 features extracted

5: 1059 features extracted

6: 1040 features extracted

7: 1022 features extracted

8: 898 features extracted

9: 1018 features extracted

10: 858 features extracted

11: 844 features extracted

12: 1035 features extracted

13: 950 features extracted

14: 983 features extracted

15: 1026 features extracted

16: 832 features extracted

17: 968 features extracted

18: 1038 features extracted

19: 1012 features extracted

20: 850 features extracted

21: 1089 features extracted

22: 985 features extracted

23: 1074 features extracted

24: 1002 features extracted

25: 1014 features extracted

26: 1083 features extracted

27: 1004 features extracted

28: 910 features extracted

29: 1007 features extracted

30: 966 features extracted

31: 1071 features extracted

32: 961 features extracted

33: 969 features extracted

34: 981 features extracted

35: 895 features extracted

36: 942 features extracted

37: 1041 features extracted

38: 1056 features extracted

39: 991 features extracted

40: 970 features extracted

41: 945 features extracted

42: 1006 features extracted

43: 975 features extracted

44: 948 features extracted

45: 858 features extracted

46: 1006 features extracted

47: 1134 features extracted

48: 1044 features extracted

49: 1013 features extracted

50: 1033 features extracted

51: 974 features extracted

52: 991 features extracted

53: 933 features extracted

54: 1001 features extracted

55: 855 features extracted

56: 1008 features extracted

57: 935 features extracted

58: 992 features extracted

59: 909 features extracted

60: 905 features extracted

61: 996 features extracted

62: 881 features extracted

63: 999 features extracted

64: 754 features extracted

65: 917 features extracted

66: 720 features extracted

67: 1006 features extracted

68: 1003 features extracted

69: 919 features extracted

70: 964 features extracted

71: 957 features extracted

72: 977 features extracted

73: 970 features extracted

74: 944 features extracted

75: 992 features extracted

76: 966 features extracted

77: 866 features extracted

78: 964 features extracted

79: 926 features extracted

80: 959 features extracted

81: 1014 features extracted

82: 990 features extracted

83: 958 features extracted

84: 926 features extracted

85: 974 features extracted

86: 953 features extracted

87: 963 features extracted

88: 987 features extracted

89: 885 features extracted

90: 846 features extracted

91: 943 features extracted

92: 926 features extracted

93: 899 features extracted

94: 883 features extracted

95: 961 features extracted

96: 957 features extracted

97: 861 features extracted

98: 902 features extracted

99: 888 features extracted

Log 2 Threads

Thread 0 started

Thread 1 started

0: 924 features extracted

1: 980 features extracted

2: 988 features extracted

3: 873 features extracted

5: 1059 features extracted

4: 985 features extracted

6: 1040 features extracted

7: 1022 features extracted

8: 898 features extracted

9: 1018 features extracted

11: 844 features extracted

10: 858 features extracted

12: 1035 features extracted

13: 950 features extracted

14: 983 features extracted

16: 832 features extracted

15: 1026 features extracted

17: 968 features extracted

18: 1038 features extracted

19: 1012 features extracted

20: 850 features extracted

21: 1089 features extracted

23: 1074 features extracted

22: 985 features extracted

24: 1002 features extracted

25: 1014 features extracted

26: 1083 features extracted

27: 1004 features extracted

28: 910 features extracted

29: 1007 features extracted

30: 966 features extracted

31: 1071 features extracted

32: 2901 features extracted

33: 969 features extracted

34: 3189 features extracted

36: 2403 features extracted

35: 895 features extracted

37: 3156 features extracted

39: 3437 features extracted

38: 1056 features extracted

40: 3045 features extracted

42: 3054 features extracted

41: 945 features extracted

43: 2716 features extracted

44: 948 features extracted

45: 2393 features extracted

47: 3039 features extracted

46: 1006 features extracted

48: 3105 features extracted

50: 3436 features extracted

49: 1013 features extracted

51: 2830 features extracted

52: 991 features extracted

53: 2945 features extracted

55: 2510 features extracted

54: 1001 features extracted

56: 3122 features extracted

58: 2941 features extracted

57: 935 features extracted

59: 3006 features extracted

60: 905 features extracted

61: 2864 features extracted

63: 3099 features extracted

62: 881 features extracted

64: 2096 features extracted

66: 1941 features extracted

65: 917 features extracted

67: 3050 features extracted

69: 2379 features extracted

68: 1003 features extracted

70: 2972 features extracted

72: 2864 features extracted

71: 957 features extracted

73: 3012 features extracted

75: 3301 features extracted

74: 944 features extracted

76: 2644 features extracted

78: 3054 features extracted

77: 866 features extracted

80: 2519 features extracted

79: 926 features extracted

81: 3177 features extracted

83: 3142 features extracted

82: 990 features extracted

84: 2637 features extracted

86: 2744 features extracted

85: 974 features extracted

87: 2934 features extracted

89: 2590 features extracted

88: 987 features extracted

90: 2962 features extracted

92: 2900 features extracted

91: 943 features extracted

93: 2878 features extracted

94: 883 features extracted

95: 2773 features extracted

97: 2392 features extracted

96: 957 features extracted

98: 2641 features extracted

100: 2738 features extracted

99: 888 features extracted

[Thomas]

I have not found a solution but this might be related to my problem (line 504 in AlignLayersTask)

// Not concurrent safe! So two copies, one per layer and Thread:

final SIFT ijSIFT1 = new SIFT( new FloatArray2DSIFT( p.sift ) );

although in the script above I was creating a new FloatArray2DSIFT(param) for each thread.

I tried also creating the FloatArray2DSIFT(param) inside the loop but the problem remained.

Thomas

[Stephan Saalfeld]

Hi Thomas,

please also clone the param object:

floatSIFT = FloatArray2DSIFT(param.clone())

...

depending on the parameters used, the parameters can be non-thread-safe.
Bad design from long time ago.

Sorry,
Stephan

> --
> --
> Please avoid top-posting, and please make sure to reply-to-all!
>
> Mailing list web interface: http://groups.google.com/group/fiji-devel
>
> ---
> You received this message because you are subscribed to the Google
> Groups "Fiji-devel" group.
> To unsubscribe from this group and stop receiving emails from it, send
> an email to fiji-devel+...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.

[Thomas]

Oh yes that was it ! It is working now.

Thank you very much

Best regards

Thomas

[Thomas]

And for the record, here is the minimalistic python script that computes in parallel the features of images contained in a folder.

Thomas

import ij, os,time

from ij import IJ

from mpicbg.ij import SIFT

from mpicbg.imagefeatures import FloatArray2DSIFT

from java.util import HashSet

from java.lang import Thread

from java.util.concurrent.atomic import AtomicInteger

from java.lang import Runtime

path = os.path.normpath(xxx)

paths = [ os.path.join(path,name) for name  in filter(lambda x: os.path.splitext(x)[1] == '.tif', os.listdir(path))]  # list containing the paths of the images

#print paths

nImages = len(paths)

atomicI = AtomicInteger(0)

threads = []

def getFeatures():

pLowRes = FloatArray2DSIFT.Param().clone()

pLowRes.fdSize = 8

pLowRes.initialSigma = 0.5

while atomicI.get() < (nImages - 1):

k = atomicI.getAndIncrement()

if k < nImages:

floatSIFT = FloatArray2DSIFT(pLowRes)

ijSIFTLowRes = SIFT(floatSIFT)

imp = IJ.openImage(paths[k])

ip = imp.getProcessor()

features = HashSet()

ijSIFTLowRes.extractFeatures(ip, features)

IJ.log( str(k) + ': ' + str(features.size()) + ' features extracted' )

imp.close()

del features

# Creating and starting the threads

for p in range(int (Runtime.getRuntime().availableProcessors() * 0.9 )):

#for p in range(1):

thread = Thread(getFeatures)

threads.append(thread)

thread.start()

#time.sleep(5)

IJ.log('Thread ' + str(p) + ' started')

# Joining the threads

for thread in threads:

thread.join()

IJ.log('Thread joined')

IJ.log('Computing features done')