Suggestion for new model type

[edward9112]

I have a suggestion for a model for quite common use case: people detection in an overhead camera scene.
Regular models like person-detection-retail just won't detect in such circumstances but this is the best camera position for non-occlusive people counting. It would be great if you add this type of model. I could assist with data collection if necessary.

Here is the image example: https://ibb.co/37Ybx5b

[vladimir-dudnik]

@snosov1 could you please review this request?

[snosov1]

Hey, @edward9112 !

We've considered such a use case previously. The closest we could get is the set of "crossroad" models that has a certain amount of "overhead" data in the training set. Though, on practice it still requires the camera to be somewhat higher up than you show on your example image.

That said, we could consider looking into it again. What kind of help in data collection do you imply?

[edward9112]

Hi @snosov1 ! I can provide video footage or assist with image annotation.
Here are some more typical examples:
https://ibb.co/mBpwYjP
https://ibb.co/8DYXH0q
https://ibb.co/WP1D8jL

[snosov1]

1. What volumes of data are we talking about (minutes, hours, days, weeks)?
2. What is the number of locations?
3. Will you be able to make this data public under some permissive terms (otherwise, getting it into our premises will surely take a while and might not even be possible in the end)?

As for the annotation - we have the means to do somewhat large-scale annotation with our resources (given that we have satisfactory answers on the questions above).

[edward9112]

1. Days, probably weeks of video data
2. Could be 5-6 plus I could try to scrape some public sources as well which should add 10-20 locations
3. Not sure about this. Does sharing it via online channel for annotation only make it public? Or you add the data to some kind of public archive?

[snosov1]

1-2. That sounds like something we can work with.

With regards to third item - the main reason I'm asking is that we, as a company, have to get data with clear terms on how we can use it. One of the simplest ways is if the dataset is public (i.e. accessible to general public, not only to us) with clear and non-restrictive terms of use/licensing. If you need to keep it between our parties - then there have to be some explicit license agreement between our companies (and if you're just you and not a company that we can have such kind of agreement with - then it's simply a no-go from our side).

[edward9112]

Do you have a draft of such agreement? How do I make a dataset public?

[snosov1]

Do you have a draft of such agreement?

No. When we purchase/acquire data - the providing company gives us its terms for review.

How do I make a dataset public?

You host the binaries on a service of your choice and have a website with the links, description, license/terms of use. Popular examples are - WIDERFace, MS COCO and many others.

[edward9112]

Alright, I will let you know as soon as I collect the required data.

[snosov1]

Thx! Looking forward to it!

[edward9112]

Thx! Looking forward to it!

Hi @snosov1

I have collected the dataset, it's around 25GB of video footage with motion in it. Many scenes and camera positions. The average resolution is 640x360px.

Would that be sufficient?

[snosov1]

Hey, @edward9112 !

Great news! The spec sounds great - we'll need to execute some experiments to see what we can achieve with it, though!

Please, refer any questions you might have with regards to sharing the data to @yssemaev . Once this is figured out and we can access the data - we'll plan the experiments accordingly.

[edward9112]

Awesome!

@yssemaev can I send you the data sharing agreement draft? If so, can you provide your contact/email?

[yssemaev]

@edward9112, please contact me directly by email: yuri.semaev at intel.com to discuss details

[edward9112]

Hi,
@yssemaev @snosov1 it looks like the agreement process is stalled. I never got any replies from your legal team.

[yssemaev]

Pinged legal team, waiting for response.

[edward9112]

@snosov1 @yssemaev it looks like the legal team's response is going to take forever.
Is there another way to pass the materials to you without putting them in a public data bank permanently?

[edward9112]

@snosov1 @yssemaev the legal team seem to have approved the agreement but then stopped responding.
Is there any way to expedite?

[edward9112]

@snosov1 @yssemaev
Hi again, just a follow-up regarding the agreement.
We spent a lot of time and resources collecting the data, so it would be unfortunate to just leave it behind.

[edward9112]

@vladimir-dudnik @eaidova
Any suggestions?

[vladimir-dudnik]

@edward9112 can you publish your data under permissive license on some public resource, like @snosov1 suggested in his comment?

[edward9112]

@snosov1 @vladimir-dudnik is it OK if we publish it on Kaggle.com ?

[vladimir-dudnik]

@edward9112 I think it should be OK if dataset is available on public resource under permissive license

[edward9112]

@snosov1 @vladimir-dudnik it looks like the agreement has been signed today. I will proceed with uploading the data for sharing.

[edward9112]

@vladimir-dudnik @snosov1 @yssemaev data has been shared with you via Google Drive

[edward9112]

@vladimir-dudnik @snosov1 @yssemaev any comments/feedback regarding data quality?

[edward9112]

@vladimir-dudnik @snosov1 @yssemaev any feedback would be helpful. Should we add more data?