4 ms·

Face Recognition Using Pytorch

119 points by timesler 7y ago

s_Hogg 7y ago
Where in the page is the evidence of the speed claim made in the headline for this submission? Edit: title has now been changed
briggers 7y ago
I see from the MTCNN code that this repo (like all others I've seen) is still bouncing tensors between GPU and CPU while passing between the P/R/ONets. So many ML repos make this mistake in pre/post-processing and end up bottlenecked on CPU. Anyone know of an MTCNN that's been ported to run more or less fully on GPU? (Or even that does batching instead of an image-by-image approach?)
- cetra3 7y ago
  I have used this before: https://github.com/blaueck/tf-mtcnn https://github.com/blaueck/tf-mtcnn which uses tensorflow for all the xNets. Example in rust: https://cetra3.github.io/blog/face-detection-with-tensorflow-rust/ https://cetra3.github.io/blog/face-detection-with-tensorflow...
- timesler 7y ago
  I'm not aware of any implementation with these features, but they are both on the roadmap for the linked repo. Both should also be achievable. Batch processing, in particular, will be a straight-forward change and should result in quite a speed-up. Although it will require the input images to have the same dimensions.
  - briggers 7y ago
    Good stuff. In my experience inputs to MTCNN tend to be full frames, so the uniform dimension requirement is usually met.
ldulcic 7y ago
Great repo, been using it for a while now! Thank You Tim!
- timesler 7y ago
  Awesome, and thanks for your feedback in the early days. I've made the tracking interface much nicer as a result.
cnxhk 7y ago
I usually use this one: https://github.com/deepinsight/insightface https://github.com/deepinsight/insightface
- timesler 7y ago
  That's a good repo. It uses mxnet right? The aim of the repo in the topic was mainly to provide a clean implementation that could slot easily into an existing pytorch workflow.
lorepieri 7y ago
Thanks for sharing. I'm working on face recognition with homomorphic encryption, therefore without compromising the user privacy. The bold goal is the first privacy preserving videocamera. If you find this interesting, I would love to chat about it.
- timesler 7y ago
  That sounds like a pretty interesting challenge. Happy to chat more - you can get in touch using the contact details on github.

119 points by timesler 7y ago

Where in the page is the evidence of the speed claim made in the headline for this submission? Edit: title has now been changed

briggers 7y ago

I see from the MTCNN code that this repo (like all others I've seen) is still bouncing tensors between GPU and CPU while passing between the P/R/ONets. So many ML repos make this mistake in pre/post-processing and end up bottlenecked on CPU. Anyone know of an MTCNN that's been ported to run more or less fully on GPU? (Or even that does batching instead of an image-by-image approach?)

cetra3 7y ago
I have used this before: https://github.com/blaueck/tf-mtcnn https://github.com/blaueck/tf-mtcnn which uses tensorflow for all the xNets. Example in rust: https://cetra3.github.io/blog/face-detection-with-tensorflow-rust/ https://cetra3.github.io/blog/face-detection-with-tensorflow...
timesler 7y ago
I'm not aware of any implementation with these features, but they are both on the roadmap for the linked repo. Both should also be achievable. Batch processing, in particular, will be a straight-forward change and should result in quite a speed-up. Although it will require the input images to have the same dimensions.
- briggers 7y ago
  Good stuff. In my experience inputs to MTCNN tend to be full frames, so the uniform dimension requirement is usually met.

ldulcic 7y ago

Great repo, been using it for a while now! Thank You Tim!

timesler 7y ago
Awesome, and thanks for your feedback in the early days. I've made the tracking interface much nicer as a result.

cnxhk 7y ago

I usually use this one: https://github.com/deepinsight/insightface https://github.com/deepinsight/insightface

timesler 7y ago
That's a good repo. It uses mxnet right? The aim of the repo in the topic was mainly to provide a clean implementation that could slot easily into an existing pytorch workflow.

lorepieri 7y ago

Thanks for sharing. I'm working on face recognition with homomorphic encryption, therefore without compromising the user privacy. The bold goal is the first privacy preserving videocamera. If you find this interesting, I would love to chat about it.

timesler 7y ago
That sounds like a pretty interesting challenge. Happy to chat more - you can get in touch using the contact details on github.