How to use the age-gender-recognition-retail-0013 ?

Will

Hi, I tested the age-gender-recognition-retail-0013 model with the demo script but this shows a very low resolution image (62 x 62 pixels) with huge text on it. The command I used:
python3 depthai_demo.py -dd -cnn age-gender-recognition-retail-0013
I then thought maybe it should be used as a 2nd stage (like emotions-recognition-retail-0003, facial-landmarks-35-adas-0002, landmarks-regression-retail-0009) but that particular model is not supported as 2nd stage.
Did I miss something? Thanks!

Brandon

Hi Will,

Thanks and sorry about the trouble. So we had actually held off on implementing age-gender support in Gen1 in lieu of supporting a more robust/flexible implementation in the Gen2 pipeline builder.

As for more general multi-stage support like this, we recommend our Gen2 pipeline builder. So it is already possible to run such multi-stage for face detection and then age/gender from that.

An example of a more complicated neural flow than this, is here:
https://github.com/luxonis/depthai-experiments/tree/master/gaze-estimation#gaze-estimation

So we have on our roadmap to implement an example with age-gender in Gen2, which would actually include multiple features in one that will do face detection and allows running the following as secondary networks run on the ROI of the face in parallel.

Age + Gender recognition
Facial Expression estimation (incorrectly called ‘emotion’ in their doc)
Facial Landmarks

But in the meantime, it is possible to use the Gen2 pipeline builder to make this flow, if you'd like. Leveraging the Gaze example above as a reference, and these more basis-function like examples below as other references.

https://github.com/luxonis/depthai-experiments/tree/master/gen2_examples

And if this flow is pressing we can prioritize making the example above - just let us know (i.e. is this a curiosity thing? Or do you need it for a pressing business need?). If the latter, and if you'd like to do it to learn the Gen2 system we are more than happy to help. Either way, please feel free to join our Discord Community to discuss real-ish time:
https://discord.gg/EPsZHkg9Nx

Thoughts?

Thanks,
Brandon

Will

Thanks for the clarification Brandon ! Nothing urgent, just exploring the APIs. I understand the Gen2 API is the way to go forward. I just tested the gaze-estimation example. It works fine with the OAK-1 but fails with the OAK-D with the following log:

python3 main.py -vid demo.mp4
Loading pipeline...
Creating pipeline...
Creating Face Detection Neural Network...
Creating Landmarks Detection Neural Network...
Creating Head Pose Neural Network...
Creating Gaze Estimation Neural Network...
Pipeline created.
Traceback (most recent call last):
File "/depthai-experiments/gaze-estimation/main.py", line 325, in <module>
Main(file=args.video).run()
File "/depthai-experiments/gaze-estimation/main.py", line 104, in init
self.start_pipeline()
File "/depthai-experiments/gaze-estimation/main.py", line 176, in start_pipeline
self.device = depthai.Device()
RuntimeError: Failed to find device after booting, error message: X_LINK_DEVICE_NOT_FOUND

The OAK-D works fine otherwise with the Gen2_examples.

On a side note, it is a bit annoying to have to re-install the DepthAI API (0.0.2.1) when testing the Gen2_examples and the other examples (eg gaze-estimation) as they seem to rely of slightly different versions. I'm aware that the Gen2 API is still a work in progress and I'm sure you are working hard on unifying the various examples under the same Gen2 API version. Thanks again !

Will

Brandon

Will Hi Will

Thanks for the report. Interesting that that fails on OAK-D but not OAK-1. Just pinged Lukasz who wrote that example to ask him to help investigate/fix.

And yes, we soon will be making the examples work all off of a core updated library to make this not a recurring annoyance. Totally understand and agree.

In the meantime, if you do a virtualenv (or conda, or similar) for each example, then they can live/coexist w/out having to reinstall each time you run each.

Thoughts?

Thanks again for the feedback!
-Brandon

Will

Hi Brandon,
Looks like Lukasz fixed the OAK-D issue with this commit: https://github.com/luxonis/depthai-experiments/pull/37
Thanks, the gaze-estimation example now works with the OAK-D !
Will

Brandon

Awesome, thanks for following up @Will (and for the fix, Lukasz and team!)