Aim Data Classification Ideas #14
Replies: 3 comments 4 replies
-
Data Collection There do exist other options... If you are willing to crowdsource video annotation. AWS Sagemaker has a tool called ground truth where we could attempt data annotation. The problem here is that there are non-trivial costs to annotating in this way. Namely, we would need to host the video somewhere, and AWS charges a fee to annotate an image through their tool, even if we are doing the annotation ourselves. More open-source data annotation tools may be applicable here, but they would require more in-depth server infrastructure in order to function. How Can this be Done? I bring up GANs here because, while we try to catch cheaters, the cheat developers ought to engage in some adversarial process to frustrate our progress. With this in mind, I would suggest that in tandem with the cheat Discriminator, that a Cheat Generator that acts in the user input space be developed in parallel. Ideally with this method, we would only be providing a cheat that the discriminator could already detect, thus it would not be a legitimate threat to a gaming ecosystem. Further in this vein of architecture selection; I would like to point out that the real user actions (i.e. keyboard presses, or joystick motion) relative to the displayed video, are what we are attempting to gauge as either human or not. It is very likely that we would need both, a representation of the video, and some extracted features for user input for this system to be useable. Conceptually this would be a comparison between user input to the game and the image data that they received. Such a mapping would likely be specific to a given game or at least a certain genre of game. |
Beta Was this translation helpful? Give feedback.
-
The first step as always is recollecting the data and check how to properly label it, also answer the following things:
Once we answer this questions/concerns we can discuss properly the machine learning algorithm (just doing a fast check it will be probably a CNN-LSTM ( like it is done for activity recognition), or an ensemble method (maybe take the images to a lower dimension with an auto encoder and train a time series model)) |
Beta Was this translation helpful? Give feedback.
-
If this project isn't dead, we could use a technique similar to that used in this article. We could generate one of these kinds of images for every ~5-10s of video and feed it into an image recognition model. Obviously the hard part is getting enough data, but a proof of concept for this wouldn't need more than an hour or so of labeled data, especially if we use a pretrained model through something like fastai. |
Beta Was this translation helpful? Give feedback.
-
Current Status
Right now, we haven't started any work on this aspect of the project.
What we Need
We need to start researching methods of classifying aim data as either cheating or human.
I've used Tensorflow/Keras in python before to do some light machine learning / neural net classification and I think it could be helpful.
Do we need to pursue a deep learning route? Is simple machine learning appropriate?
How should we store and tag the aiming data to feed our program?
What variables do we need from the aiming data?
Post any ideas, research, or thoughts here!
Beta Was this translation helpful? Give feedback.
All reactions