aai-project

Motivation

Dataset

Original Dataset

https://www.kaggle.com/datasets/tr1gg3rtrash/yoga-posture-dataset

The dataset contains 2,700 + images classified into 46 categories.

Cleaned Dataset

Thanks to Lukas’s work, some obvious outliers that were classified into the wrong category were removed. Then we get a cleaner dataset with 2300+ images and 47 classes.

Enhanced Dataset

Thanks to the work of Lukas and Alex, more yoga posture images are added. At the same time, some underrepresented classes are merged, and new poses​ are added. The dataset increases from 2300 to 4000 images, representing 46 classes.

Approach

• Image Classification
• Usage of Convolutional Neural Networks & Vision Transformer​
• Tried both pretrained and non-pretrained models for accuracy optimization

1. CNN

2. Train ViT From Scrach

3. ViT Based Fine Tuning

Validation accuracy on original dataset(2700+ images, 47 classes)

***** eval metrics *****
  epoch                   =        4.0
  eval_accuracy           =     0.9275
  eval_loss               =     0.4354
  eval_runtime            = 0:00:27.82
  eval_samples_per_second =     14.881
  eval_steps_per_second   =      1.869

Validation accuracy on cleaned dataset(2300+ images, 47 classes)

***** eval metrics *****
  epoch                   =        4.0
  eval_accuracy           =     0.9528
  eval_loss               =     0.3528
  eval_runtime            = 0:00:23.16
  eval_samples_per_second =     14.636
  eval_steps_per_second   =      1.856

Validation accuracy on enhanced dataset(4000 images, 46 classes)

***** eval metrics *****
  epoch                   =        4.0
  eval_accuracy           =     0.9754
  eval_loss               =     0.1424
  eval_runtime            = 0:01:05.76
  eval_samples_per_second =      8.668
  eval_steps_per_second   =      1.095

Performance Comparison

Conclusion