Specified models for pedestrian_tracker_demo

[dkustikx]

Specified list of supported models.
Models are not supported by this demo: person-detection-retail-0002, person-reidentification-retail-0076, person-reidentification-retail-0079.

[openvino-pushbot]

Can one of the admins verify this patch?

[openvino-pushbot]

Can one of the admins verify this patch?

[openvino-pushbot]

Can one of the admins verify this patch?

[openvino-pushbot]

Can one of the admins verify this patch?

[openvino-pushbot]

Can one of the admins verify this patch?

[vladimir-dudnik]

Jenkins please retry a build

[kchechil]

cc @IRDonch

[IRDonch]

Sounds like a bug in the demo to me.

[IRDonch]

@Wovchena So would it be possible to update pedestrian_tracker_demo to make it work with the other model variants? From the descriptions, it doesn't sound like they're that different.

[kchechil]

we need to think how to align the demo instead of narrowing down the list of supported models.

[Wovchena]

• I am suggesting to keep ReID models consistent. Currently person-reidentification-retail-0079 and person-reidentification-retail-0076 have output shape of [1, 256, 1, 1], while person-reidentification-retail-0031 - [1, 256], but person-reidentification-retail-0031.md claims it's [1, 256, 1, 1]. I think [1, 256] makes more sense. The demo expects [1, 256].
• There is an update for detection part: pedestrian_tracker: generalize detection #286

[kchechil]

@snosov1 @Ilya-Krylov is it possible to align all reid models to the same output format?

[snosov1]

@evgeny-izutov the difference in output format is, probably, an accident?

[evgeny-izutov]

The person-reidentification-retail-0031 model uses InnerProduct layer as final layer, so the output resolution is [1, 256]. The rest models are converted to support 1x1 conv output layer (output resolution is [1, 256, 1, 1]). It looks like the demo use the old variant of IE API to determine the embedding size (256) assuming the reversed order of shape operation.
P.S. I've never seen this demo before.

[snosov1]

@Ilya-Krylov can we get away with the changes only on the demo side?

[Ilya-Krylov]

@snosov1 @kchechil I think it is possible, some kind of reshape operation should be added. I would propose to do this to @LeonidBeynenson as author of the demo.

[snosov1]

Marvelous idea! @LeonidBeynenson what do you think?

[LeonidBeynenson]

I created pull request #295
It contains the changes to support the reid models person-reidentification-retail-0076 and person-reidentification-retail-0079.

Please, note that the PR #295 contains models.lst file too, so if #295 is merged this PR may be closed as a duplication.

[LeonidBeynenson]

Sorry, after small talk with @Wovchena I removed my changes from models.lst file, since now (after my PR #295) it supports all reid models and, as far as I understand, after changes by @Wovchena in #286 it supports all person detection models.

[LeonidBeynenson]

Please, note that, if I understood @Wovchena correctly, the change in models.lst file for pedestrian tracker demo (i.e. this PR) will not be required after merging @Wovchena PR #286 and my PR #295.

[IRDonch]

Closing, since this is obsoleted by the aforementioned PRs.