perfmance: "imageDatastore" low performance when specifying multiple file names directly

Very low performance when specifying multiple image full path names for a given imageDatastore，see benchmark below.

env: matlab 2022b， win10

imgsDir = "someImgDir";
imgNames = imgsDir+"/"+string(0:2687)+".jpg";% The more images you have, the lower the performance

t1 = tic;
for i = 1:10
    imds = imageDatastore(imgNames); % low
end
t = toc(t1)

t2 = tic;
for i = 1:10
    imds = imageDatastore(imgsDir); % high
end
t = toc(t2)

t =

14.0975

t =

0.5759

Initially I wanted to sort the files in the ImageDatastore by Arabic arrays, but found the above to be a real performance problem, any better suggestions?

Answer 1

[Rahul]

Hi @cui,xingxing,

I understand that reading images while specifying the image names is causing performance issues as compared to reading directly from a folder. However, reading from a folder would limit the ability to sort it before reading.

Here are a few suggestions which could improve the performance of your use-case:

• Sort the images after reading them. This can be done by extracting the filenames and then sorting them. Here is an example:

imds = imageDatastore(imgsDir);

% Extract file names and sort numerically
[~, name, ext] = cellfun(@fileparts, imds.Files, 'UniformOutput', false);
nums = str2double(name);
[~, idx] = sort(nums);

% Reorder imds.Files
imds.Files = imds.Files(idx);

• Read images using Batching or Lazy-Loading techniques.

The following MATLAB Answer might also be useful:

https://www.mathworks.com/matlabcentral/answers/382692-how-to-optimize-imagedatastore-speed

The following MathWorks documentations can be referred:

'imageDatastore': https://www.mathworks.com/help/matlab/ref/matlab.io.datastore.imagedatastore.html

'fileDatastore': https://www.mathworks.com/help/matlab/ref/matlab.io.datastore.filedatastore.html

'augmentedDatastore': https://www.mathworks.com/help/deeplearning/ref/augmentedimagedatastore.html

Thanks.