Thank you for your presentation. According to your paper, you trained your model by using 3d Convnet (MIL) to get the highest score data, not by resorting to human annotators. Could you share the detailed process how to calculate and predict (when it is predicted value, meaning you have to have some data annotated by human for traning?) the highest scores from 3d Convnet based on the MIL technique?
I am also having doubts since they trained it on a per video segment level but the metric used was a frame-based ROC. I'm planning to contact the author via mail and clarify this.
@@jessiejamessuarez8446 Two major doubts I have after implementing the method in this paper: 1. The anomaly scores output for anomaly segments in testing anomaly videos are not actually close to 1. In fact, they vary from around 10^-6 to 1. And the threshold i got when drawing the ROC curve is mostly around 10^-5. This means that the anomaly boundary learned by the method is not 0.5 but some other number quite small. 2. If you test its ROC only on anomaly videos you will find that the AUC drops badly from 0.7X to 0.5X. I visualized the scores of the anomaly videos as the authors did and found that most anomaly curves are actually NOT as 'beautiful' as the authors showed us
@@davidyang9191 Hello, please I want to implement this project but I don't know how to get started. Can you recommend any resources or share your code? Thank you for your time.
Thank you for your presentation. According to your paper, you trained your model by using 3d Convnet (MIL) to get the highest score data, not by resorting to human annotators. Could you share the detailed process how to calculate and predict (when it is predicted value, meaning you have to have some data annotated by human for traning?) the highest scores from 3d Convnet based on the MIL technique?
Thankyou, I've learned a lot from your paper!
thank you for explaining clearly how it works.
How to down load code and datasets.,, and how execute the code on python
no code for the paper ?
Hi Dear Chen
How can you to Create animate plot?
Anyone understood this paper I have a doubt in it
I am also having doubts since they trained it on a per video segment level but the metric used was a frame-based ROC. I'm planning to contact the author via mail and clarify this.
@@jessiejamessuarez8446 Two major doubts I have after implementing the method in this paper:
1. The anomaly scores output for anomaly segments in testing anomaly videos are not actually close to 1. In fact, they vary from around 10^-6 to 1. And the threshold i got when drawing the ROC curve is mostly around 10^-5. This means that the anomaly boundary learned by the method is not 0.5 but some other number quite small.
2. If you test its ROC only on anomaly videos you will find that the AUC drops badly from 0.7X to 0.5X. I visualized the scores of the anomaly videos as the authors did and found that most anomaly curves are actually NOT as 'beautiful' as the authors showed us
@@davidyang9191 Hello, please I want to implement this project but I don't know how to get started. Can you recommend any resources or share your code? Thank you for your time.
May I have full text ?
What are you bubbling about?
any code of anomaly detection?
Bro I want to purchase the project...how to contact you
U probably wouldn't like me much I'm lookin for ghosts
why is f(Va) is the most likely true positive instance ? Isnt f is just a function that we invent ?
very poor explanation