Experience with 32-Node Model or Similar Models Focusing on an Animal's Legs

[pelikanGIS]

Hi everyone,

I’m currently working on building a model that focuses on tracking an animal's legs, and I’m exploring the idea of using a 32-node model or a similar configuration. While reviewing the SLEAP datasets documentation, I noticed limited discussion about this approach. The only reference I found to the 32-node model in the forum is from this thread.

I am part of the Deutsch Lab at the University of Haifa, and I’m looking to refine my approach to building a robust model for this purpose.

If you’ve attempted to apply the 32-node model or a similar one with an emphasis on tracking an animal's legs, I’d love to learn about your experience. Were there any specific challenges, modifications, or insights that stood out? Any guidance or advice would be incredibly helpful as I work to improve my model.

Thank you!

Best regards,
Ehud Sussman
Deutschlab, University of Haifa

[eberrigan]

Hi @pelikanGIS,

Please check out these discussions and let us know if you have any questions! #1977 #1998

I believe you will find the second one especially helpful.

Best,

Elizabeth

[talmo]

I'll just add: increasing the number of filters or filters rate, or decreasing the output stride) will be important for high node counts.

The reason for this is that having too many nodes will conflict with a bottleneck in the penultimate layer prior to the confidence maps depending on how it's set up.

The equation for the number of filters in the penultimate layer is filters * (filters_rate ** (1 - output_stride)).

For example, let's say this is our configuration:

• filters = 16
• filter_rate = 0.5
• output_stride = 2

Then the penultimate layer will have 16 * (0.5 ** (1 - 2)) = 32 channels.

If we set the output_stride = 1, however, we get 16 * (0.5 ** (1 - 1)) = 16 channels in the penultimate layer.

If you have a layer with 16 channels that then outputs to a layer that has 32 channels (= nodes in the confidence maps), it won't have sufficient representational capacity to properly regress all the nodes, and will end up underperforming on some of them.

Let us know if you have any questions!

Cheers,

Talmo