CODEWITHSUNDEEP
pythonopencvorbfeature-descriptor
drheinrich940
drheinrich940

Reputation: 143

Why is it faster to recompute ORB descriptor than loading it from disk?

I am trying to decide whether or not it is interesting to persist inside local files or inside a database the computed descriptors of a huge amount of pictures ( each .png picture has a resolution of 500x500, and weights aproximatively 25kb ).

Using ORB with Brief-32 descriptors, a single descriptor weights around 3 megabytes. Such size would remain constant as all my pictures are of the same dimensions.

To find out what was the fastest i ran the two following tests :

## TEST : Import descriptor from file
listOfDec = list()
start = datetime.now()

for i in range(0, 100):
    listOfDec.append(np.loadtxt("DESC_TEST".txt"))

end = datetime.now()
time_taken = end - start
print('Time: ',time_taken) 

## TEST : Compute descriptor from source image
listOfDec = list()
start = datetime.now()

for i in range(0, 100):
    img1 = cv2.imread(dirPath+picture,0)
    a, desc = orb.detectAndCompute(img1, None)
    listOfDec.append(desc)

end = datetime.now()
time_taken = end - start
print('Time: ',time_taken)

I honestly thought it would be faster to load the data than to recalculate the whole descriptor.

Here is the result of my test :

benchmark plot

So now i am confused. I know that ORB is a really fast algorithm, but how is it faster to "generate" a 3MB descriptor than to read it from an ssd disk ? Is there something wrong with my benchmark ?

Thank you.

Upvotes: 2

Views: 322

Answers (1)

varankou
varankou

Reputation: 901

RAM memory is really much faster than disk memory.

However 0.08sec for 3mb file gives read speed of 37.5mb/s which looks to low for SSD. And you should mention size of loaded images since ORB computations on your benchmark include reading them from disk.

ORB descriptor is very fast itself since computations are trivial and image data may be well cached in CPU. However to achieve better performance you should try C++ implementation instead of python. And there may not be a reason to store full descriptors. Common practice is to save hashes or descriptors parts depends on your use case (but full desrciptors may be needed too in some cases).

Upvotes: 2

Related Questions