desert0009 / unsupervised_filter_data

Filter and collect data via unsupervised learning

Motivation

Is possible to collect/filter useful data via unsupervised learning approach to save manual checking time?

Pipeline

Cluster the Yoga crawler mess data to collect grouping data

Sample Result

Data Visualization - t-SNE

The Seated Yoga Poses group together in the upper middle part of the picture.

TODO: 
Try to use pose-estimation instead of ImageNet pretrained model to extract people embeddings

Clustering samples - k-means

Filter out the outlier(fused with black image) with larger cosine distance

TODO: 
Try to use Mask-RCNN to crop people to remove background noise

• 

• 

• 

How to run

• In Tensorflow2.x

• Collect the mess data

  • Sample Yoga Crawler Mess Poses

$ python main.py --data_path your_crawler_mess_poses

Conclusion

Maybe we can try to use segmentation to remove people background noise and use pose-estimation to extract good embeddings to get better results in the future.