Model accuracy for tflite pose detection using a custom classifier

Hi I’ve been trying to train tflite pose classifiers to work with the movenet pi example

I’ve used the suggested tutorial and colab

and carefully followed the example yoga pose dataset (with train and test examples) but the classification results seem almost random (2 classifiers out 4 successfully classify a pose about 50% of the time).

My dataset is here

The training in the colab works well with a confusion matrix with no errors. The model accuracy is also high (0.9977) and the increasing accuracy through the training epochs looks as expected. I wondered if anyone has used this colab with success?

Thanks!

2 Likes

Hi, @rod_Dickinson

I apologize for the delayed response, I do understand that model accuracy in python environment is high as compared to Raspberry Pi environment and there might be below reasons

Training data might not represent real-world scenarios on the Raspberry Pi. Gather more diverse images covering various angles, lighting conditions and variations in pose execution and apply data augmentation techniques like rotation, flipping, and cropping to increase the dataset size and improve generalization.

Raspberry Pi hardware might be less powerful than your development environment leading to differences in computational precision and performance.

Please explore techniques like model pruning, quantization or knowledge distillation to reduce model size and computational requirements. Try different quantization techniques, please refer this official documentation for model optimization

If your Raspberry Pi supports hardware acceleration (e.g GPU or TPU) enable it to improve performance.

If there is discrepancies in input data preprocessing between the python environment and Raspberry Pi can impact accuracy. double check that preprocessing steps (e.g. normalization, resizing) are identical in both environments.

If I have missed something here please let me know.

Thank you for your cooperation and patience.