Rekognition Custom Model Training with true negative images

0

I am trying to train the custom labels model using a dataset. Since the bounding boxes are mandatory to train the model, i need to remove some of the true negative images which does not have any labels. Due to this I believe that the performance of the model is degraded a little hampering to reduce the precision. I want to understand how can i add the true negative images without any labels for the training of the model. Also correct me if my understanding is not correct with the model training.

已提問 4 個月前檢視次數 148 次
2 個答案
0

Certainly! To improve your Amazon Rekognition Custom Labels model:

  • Define the attributes you want to identify.
  • Collect relevant data.
  • Add negative labels for better accuracy.

You’re on the right track!

profile picture
專家
已回答 4 個月前
profile pictureAWS
專家
已審閱 4 個月前
  • When you mention negative labels, what should i exactly do here. I do not understand the last bullet point. Thanks for the help

0

This blog post covers negative labels with Rekognition.
https://aws.amazon.com/blogs/machine-learning/tips-to-improve-your-amazon-rekognition-custom-labels-model/

Hope this helps!

profile pictureAWS
專家
已回答 4 個月前
  • Thanks for the response. I have gone through the link you shared. I still have doubt regarding this. I am using a bounding box for detecting an object. When you say the negative label then for such an image what bounding box shall i provide? This creates another scenario for me. If i provide the classification label instead of bounding box detection then do i also need to provide the true class classification label for training the custom model?

您尚未登入。 登入 去張貼答案。

一個好的回答可以清楚地回答問題並提供建設性的意見回饋,同時有助於提問者的專業成長。

回答問題指南