Fix NN resize with dimentions > 4 by keeper121 · Pull Request #16497 · opencv/opencv

keeper121 · 2020-02-03T23:04:51Z

This pullrequest changes

This resolves issue #15075
Fixes NN resize image last dimension values with odd dims > 4.

force_builders=linux,docs

asmorkalov · 2020-02-04T05:06:38Z

@keeper121 Thanks a lot for contribution! Could you add test for the case you've fixed?

keeper121 · 2020-02-07T12:24:44Z

@asmorkalov Hi, what format the test case should look like? This bug comes from #15075.

It has python script which do nearest neighbor resize.

import numpy as np
import cv2

# up to DIM=4 is OK
DIM = 5
T = np.zeros((12, 12,DIM)).astype(np.uint8)
T[2, 2, DIM-1] = 1
print("Sum: ", T.sum())
print(T[:, :, DIM-1])
R = cv2.resize(T, None, fx=0.9, fy=1,interpolation=cv2.INTER_NEAREST)
# it does not help to supply a pre-allocated array either
#R = np.zeros((12, 11, DIM)).astype(np.uint8)
#R = cv2.resize(T, dsize=R.shape[:2], dst=R, interpolation=cv2.INTER_NEAREST)

print("Sum: ", R.sum())
print(R[:, :, DIM-1])

Output before:

Sum:  1
[[0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 1 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]]
Sum:  836
[[248 113   0 214 176  85   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]
 [  0   0   0   0   0   0   0   0   0   0   0]]

Output after fix:

Sum:  1
[[0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 1 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0 0]]
Sum: 1
[[0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]
 [0 0 1 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]
 [0 0 0 0 0 0 0 0 0 0 0]]

The PR fixes case when input image has odd size of channel dimension more than 4. It occurs in https://github.com/keeper121/opencv/blob/57c99914133bea42b1fd6e176ebf35e210d7163d/modules/imgproc/src/resize.cpp#L1044-L1050 where the last image's channel part is not going through the loop.

asmorkalov · 2020-02-11T07:04:35Z

c++ test with the test data you posted before is enough.

asmorkalov · 2020-02-11T07:06:26Z

Also you can extend existing C++ test for resize to case with 4 channels too.

alalek · 2020-02-11T10:06:20Z

extend existing C++ test for resize

Separate test is better.
(existing C++ test based on legacy code which generates unstable input data)

keeper121 · 2020-02-13T19:02:06Z

@alalek Hi, I added separate test.

I run it with:

opencv_test_imgproc --gtest_filter=Resize_NN.accuracy

output before fix

Note: Google Test filter = Resize_NN.accuracy
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from Resize_NN
[ RUN      ] Resize_NN.accuracy
/media/keeper121/seagate/opencv/modules/imgproc/test/test_resize_nn.cpp:25: Failure
Expected equality of these values:
  i
    Which is: 5
  sum
    Which is: 6773
Resize NN TEST mat from 12x12x5 failed with sum 6773 should be 5
/media/keeper121/seagate/opencv/modules/imgproc/test/test_resize_nn.cpp:25: Failure
Expected equality of these values:
  i
    Which is: 7
  sum
    Which is: 30789
Resize NN TEST mat from 12x12x7 failed with sum 30789 should be 7
/media/keeper121/seagate/opencv/modules/imgproc/test/test_resize_nn.cpp:25: Failure
Expected equality of these values:
  i
    Which is: 9
  sum
    Which is: 7515
Resize NN TEST mat from 12x12x9 failed with sum 7515 should be 9
[  FAILED  ] Resize_NN.accuracy (0 ms)
[----------] 1 test from Resize_NN (0 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test case ran. (0 ms total)
[  PASSED  ] 0 tests.
[  FAILED  ] 1 test, listed below:
[  FAILED  ] Resize_NN.accuracy

 1 FAILED TEST

output after fix

Note: Google Test filter = Resize_NN.accuracy
[==========] Running 1 test from 1 test case.
[----------] Global test environment set-up.
[----------] 1 test from Resize_NN
[ RUN      ] Resize_NN.accuracy
[       OK ] Resize_NN.accuracy (0 ms)
[----------] 1 test from Resize_NN (0 ms total)

[----------] Global test environment tear-down
[==========] 1 test from 1 test case ran. (0 ms total)
[  PASSED  ] 1 test.

Should I add something else?

asmorkalov · 2020-02-14T06:24:59Z

There are build warnings on Windows:

C:\build\precommit_windows32\opencv\modules\imgproc\test\test_resize_nn.cpp(13): warning C4305: 'initializing': truncation from 'double' to 'float' [C:\build\precommit_windows32\build\modules\imgproc\opencv_test_imgproc.vcxproj]
C:\build\precommit_windows32\opencv\modules\imgproc\test\test_resize_nn.cpp(24): warning C4244: 'initializing': conversion from 'double' to 'float', possible loss of data [C:\build\precommit_windows32\build\modules\imgproc\opencv_test_imgproc.vcxproj]

Please take a look and fix.

asmorkalov · 2020-02-14T06:37:18Z

@keeper121 I checked branch 3.4 and the issue is reproducible there too. Could you rebease it on top of 3.4 and we will merge the changes from 3.4 into master during regular procedure (weekly/bi-weekly).

So, please:

change "base" branch of this PR: master => 3.4 (use "Edit" button near PR title)
rebase your commits from master onto 3.4 branch. For example:
git rebase -i --onto upstream/3.4 upstream/master
(check list of your commits, save and quit (Esc + "wq" + Enter)
where upstream is configured by following this GitHub guide and fetched (git fetch upstream).
push rebased commits into source branch of your fork (with --force option)

Note: no need to re-open PR, apply changes "inplace".

alalek

Thank you for contribution!

alalek · 2020-02-14T11:49:47Z

modules/imgproc/test/test_resize_nn.cpp

+namespace opencv_test { namespace {
+
+        TEST(Resize_NN, accuracy)
+        {


No need to create a new file.
Please put code after this resize test: https://github.com/opencv/opencv/blob/4.2.0/modules/imgproc/test/test_imgwarp.cpp#L1403-L1414

Please add issue number into test name.

Please avoid indentation in namespaces.

alalek · 2020-02-14T11:51:01Z

modules/imgproc/test/test_resize_nn.cpp

+            const int N = 10;
+            double scale_factor = 0.9;
+            int i1 = 5, j1 = 5; // position of pixel
+            for (int i = 2; i < N; ++i)


Please test 5 channels only.
Lets avoid overcomplicated tests.

alalek · 2020-02-14T11:55:37Z

modules/imgproc/test/test_resize_nn.cpp

+                dst = dst.reshape(1, dst.cols * dst.rows * dst.channels());
+                double sum = cv::sum(dst)[0];
+                EXPECT_EQ(i, sum) << "Resize NN TEST mat from " << cols << "x" << rows << "x" << i << " failed with sum " << sum << " should be " << i;


3 lines alternative:

EXPECT_EQ(i, cvtest::norm(dst, NORM_L1)) << test.size;`

…ls test only.

alalek

Well done! Thank you for contribution 👍

alalek · 2020-02-17T13:01:20Z

Fix has been reverted: #16600

Nightly failure example.

keeper121 · 2020-02-17T14:37:09Z

@alalek How can I debug it?

valgrind-3.15.0 on my Ubuntu 19.10 (g++ 7.5) is not showing any error. I'll try to fix it if it possible.

* Fix NN resize with dimentions > 4 * add test check for nn resize with channels > 4 * Change types from float to double * Del unnecessary test file. Move nn test to test_imgwarp. Add 5 channels test only.

keeper121 mentioned this pull request Feb 3, 2020

Python - cv2.resize with INTER_NEAREST return corrupted slice for channels > 4 #15075

Closed

asmorkalov added category: imgproc and removed pr: needs test New functionality requires minimal tests set labels Feb 14, 2020

asmorkalov added the pr: needs rebase Rebase patch (and squash fixup commits) on the top of target branch label Feb 14, 2020

keeper121 changed the base branch from master to 3.4 February 14, 2020 08:40

Vladimir Mikhelev added 3 commits February 14, 2020 11:42

Fix NN resize with dimentions > 4

2967421

add test check for nn resize with channels > 4

3776368

Change types from float to double

86697c7

keeper121 force-pushed the master branch from 62a5ad5 to 86697c7 Compare February 14, 2020 08:43

alalek reviewed Feb 14, 2020

View reviewed changes

Del unnecessary test file. Move nn test to test_imgwarp. Add 5 channe…

68fc433

…ls test only.

alalek approved these changes Feb 16, 2020

View reviewed changes

alalek merged commit d84360e into opencv:3.4 Feb 16, 2020

alalek mentioned this pull request Feb 17, 2020

imgproc: revert resize changes from PR 16497 #16600

Merged

alalek mentioned this pull request Feb 20, 2020

Merge 3.4 #16627

Merged

alalek mentioned this pull request Mar 29, 2020

imgproc(resize): fix resizeNNInvoker handling of generic pixel size #16928

Merged

Uh oh!

Conversation

keeper121 commented Feb 3, 2020 • edited by alalek Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

This pullrequest changes

Uh oh!

asmorkalov commented Feb 4, 2020

Uh oh!

keeper121 commented Feb 7, 2020

Uh oh!

asmorkalov commented Feb 11, 2020

Uh oh!

asmorkalov commented Feb 11, 2020

Uh oh!

alalek commented Feb 11, 2020

Uh oh!

keeper121 commented Feb 13, 2020

Uh oh!

asmorkalov commented Feb 14, 2020

Uh oh!

asmorkalov commented Feb 14, 2020

Uh oh!

alalek left a comment

Choose a reason for hiding this comment

Uh oh!

alalek Feb 14, 2020

Choose a reason for hiding this comment

Uh oh!

alalek Feb 14, 2020

Choose a reason for hiding this comment

Uh oh!

alalek Feb 14, 2020

Choose a reason for hiding this comment

Uh oh!

alalek left a comment

Choose a reason for hiding this comment

Uh oh!

alalek commented Feb 17, 2020

Uh oh!

keeper121 commented Feb 17, 2020

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

keeper121 commented Feb 3, 2020 •

edited by alalek

Loading