123 points by codeslinger123 6 months ago flag hide 13 comments
deeplearn 6 months ago next
Fascinating progress in sign language recognition! This could help so many people in the Deaf community. I'm impressed with the accuracy and real-time abilities of these deep learning models. What library/framework did they use for implementation?
pytorchmaster 6 months ago next
@deeplearn: They used PyTorch with some custom modifications to the existing ConvLSTM and CTC modules. They've released a part of the codebase for research purposes!
astuteguy 6 months ago next
@pytorchmaster: Thank you for the insights. I've got a problem related to model training and might need your help. Will DM you for more info.
tensorfreak 6 months ago prev next
@pytorchmaster: Very nice! Any loss function they used in addition to CTC? I'm working on something similar and would appreciate any tips. :)
pytorchmaster 6 months ago next
@tensorfreak: They used CTC loss with an additional time-reversal data augmentation technique. You could try implementing it and conduct a quick A/B test to see if it works for your dataset.
tensorfreak 6 months ago next
@pytorchmaster: I'm going to give it a try and report back with the results. Thanks for your help!
signer 6 months ago prev next
This gives me hope that technology can finally bridge the gap between the deaf community and non-signers. There's still much work to do, but we're on the right path. Thanks for sharing!
signer 6 months ago next
@deeplearn: I completely agree! Real-time, accurate systems can make a huge difference in education and workplace settings. Thank you for your work!
mlbeginner 6 months ago prev next
This is a great example of using deep learning to solve day-to-day issues. Recently, I've started getting into machine learning via fast.ai. The journey has just begun!
helpinghand 6 months ago next
@mlbeginner: That's awesome! fast.ai is a fantastic place to start. If you've got a question or need help with a concept, don't hesitate to ask online communities for help. Good luck on your machine learning journey!
csprof 6 months ago prev next
@mlbeginner: Sign language recognition models can lead the way in uncovering more real-life applications in computer vision and NLP. Keep learning and experimenting with different ML concepts!
cvlstmexpert 6 months ago prev next
The use of ConvLSTMs here is really interesting for sequence prediction. What were the sequence lengths and how did you handle variable length sequences during training?
signmanager 6 months ago next
@cvlstmexpert: They used bucketing with variable length sequences. The maximum sequence length was set to 120 frames after experimenting with different lengths and impacts on training time vs accuracy.