search

anibali / margipose

3D monocular human pose estimation
Apache License 2.0
98 stars 20 forks source link

Looking for a speed up #14

Closed jaggernaut007 closed 4 years ago

jaggernaut007 commented 4 years ago

Margipose is quite an accurate model, does some great detections! Kudos to the team!! I am looking to speed up the model, for fast inference. Its currently running at a speed of 2-3 FPS for inference on my 1080. The best way to speed up I have in mind would be to reduce the res of input and use the resnet 18 model. Resnet 18 wouldnt make a lot of difference keeping in mind the model architechture. Is there any other way? Thanks!

anibali commented 4 years ago

It's also possible to simply reduce the number of stacks. The pretrained models I've made available have 4 stacks, but you could try reducing that to maybe 2 and see if that reaches a better accuracy/speed trade-off.

If you do try ResNet-18 or find some other sweet spot for performance, please report back here---I'd love to hear your results.

jaggernaut007 commented 4 years ago

Yup, the stacks can be reduced. For a baseline, let me move to resnet18 and reduce stack. Lots of performance gain strategies to try out! Ill update the results here.

guker commented 4 years ago

Yup, the stacks can be reduced. For a baseline, let me move to resnet18 and reduce stack. Lots of performance gain strategies to try out! Ill update the results here.

@jaggernaut007 how much faster than original margipose using your resnet18?

Tomiha commented 4 years ago

Yup, the stacks can be reduced. For a baseline, let me move to resnet18 and reduce stack. Lots of performance gain strategies to try out! Ill update the results here.

Same question as @guker :)