[OTE][Release][XAI] Detection fix two stage bbox_head error

[dongkwan-kim01]

This is the fix of the error issued in issue, which fixes the simple_test function in the classSAMDetectorMixin.
And resolving the tiling detection length Error regarding the output size assertion.

[sungmanc]

run ote-test

[sungmanc]

Check the results at first: https://ci-ote.iotg.sclab.intel.com/job/ote/job/pr-ote-test/809/

[dongkwan-kim01]

run ote-test

[negvet]

Are you sure that it fails because of the number of saliency maps?

Basically saliency map tensor created for every dataset point. It has dimensions [N, H, W]. Where N is the total number of classes for torch task or number of predicted classes for the OV task.

But the number of saliency maps should be equal to len(dataset) in any case I believe

[sungmanc]

Are you sure that it fails because of the number of saliency maps?

Basically saliency map tensor created for every dataset point. It has dimensions [N, H, W]. Where N is the total number of classes for torch task or number of predicted classes for the OV task.

But the number of saliency maps should be equal to len(dataset) in any case I believe

Then, could you check this FAIL case? (http://validationreports.sclab.intel.com:8006/reports/test_result/638e0a31fd6d37c8e261ba3a)

The error log was as below:

The same errors were occurred in detection + tiling cases. @eugene123tw do you have any idea for this issue?

[sungmanc]

Test results: http://validationreports.sclab.intel.com:8006/reports/build_number_report?test_session_build_number=pr-ote-test-810&environment=iotg-dev-workstation-08

Most FAIL cases were solved. Although there were still two FAIL cases, it was not related to this PR. How about merging MPA first and this PR to OTE? @harimkang @dongkwan-kim01 @goodsong81 @negvet

[harimkang]

Checking the test results, one is a score drop in cls (train vs export) problem, and one is OOM. I think the existing issue appears to have been resolved.

I have confirmed this

[goodsong81]

LGTM. I've approved MPA PRs.
@sungmanc Could you merge MPA PRs?
@dongkwan-kim01 Could you update MPA commits after merge?

[dongkwan-kim01]

I've talked with Evgeny, but in the tile detection case, it seems like many saliency maps are exported for each tile.
So exception handle which saliency_maps = [saliency_maps[i] for i in range(dataset.num_samples)], existed in the previous develop, can avoid the assertion error but maybe not make the appropriate saliency map(maybe saliency maps of first few tiles).
Maybe saliency maps merging function or getting the saliency map from the original image(before tiling) would be needed.

LGTM. I've approved MPA PRs. @sungmanc Could you merge MPA PRs? @dongkwan-kim01 Could you update MPA commits after merge?

Yes, I'll update after the merge but I think we need some fixes in the saliency map for tiling case.. Maybe @negvet investigating and make another PR for it.

[negvet]

run ote-test

[negvet]

I've talked with Evgeny, but in the tile detection case, it seems like many saliency maps are exported for each tile. So exception handle which saliency_maps = [saliency_maps[i] for i in range(dataset.num_samples)], existed in the previous develop, can avoid the assertion error but maybe not make the appropriate saliency map(maybe saliency maps of first few tiles). Maybe saliency maps merging function or getting the saliency map from the original image(before tiling) would be needed.
Yes, I'll update after the merge but I think we need some fixes in the saliency map for tiling case.. Maybe @negvet investigating and make another PR for it.

Actually it works just fine. Not that good, but functionally correct.
The first tile is the full image size - so saliency_maps = [saliency_maps[i] for i in range(dataset.num_samples)] actually provide the saliency maps for the whole images, which is correct (I deleted this line of code accidently, @dongkwan-kim01 correctly get it back).

The other problem here is that the model is not able to distinguish objects at full resolution (after being trained at tiles):

But it still can see the object occurrences (you can see the difference in object density at saliency maps) and can distinguish the background (it is SSD model)

It would be good to revisit this and merge saliency maps from tiles instead of generating it from full-res image. But for now it is fine IMO.

[wonjuleee]

LGTM

[dongkwan-kim01]

@negvet Thank you! So the original one was correct. Thank you for your confirmation.

[dongkwan-kim01]

run ote-test

[dongkwan-kim01]

Since there some code change regarding Evgeny's commit and assertion revert, I ran the tile-detection in tests/ote_cli/external/model-preparation-algorithm/ote_cli/test_tiling_detection.py it locally and passed but also running the ote-test.

[sungmanc]

Test results: http://validationreports.sclab.intel.com:8006/reports/build_number_report?test_session_build_number=pr-ote-test-815&environment=iotg-dev-workstation-10

The base branch was changed.