Lecture 27: Inverse Graphics/Research In Angjoo's Group (64)
wcyjames
To predict on occluded body parts, would it largely depend on the training data? I am wondering if there are more scenes that people work while standing in the training data, then will the model be more likely to predict standing instead of sitting?
ZizhengTai
@wcyjames, that's totally possible, and in general what the predictions look like for things with occlusion largely depends on training data. What I'm more curious about it, assuming the training data contains roughly an equal amount of standing/sitting/lying positions, then will the prediction be one of them with some probability, or could it be a monstrous mixture of all of them?
selinafeng
This seems to demonstrate the point from a few slides ago that one of the challenges of this kind of prediction is how data hungry it is. The training data might need a lot of data points with occluded legs and labels that tell the model what the legs should look like when they're obscured.
To predict on occluded body parts, would it largely depend on the training data? I am wondering if there are more scenes that people work while standing in the training data, then will the model be more likely to predict standing instead of sitting?
@wcyjames, that's totally possible, and in general what the predictions look like for things with occlusion largely depends on training data. What I'm more curious about it, assuming the training data contains roughly an equal amount of standing/sitting/lying positions, then will the prediction be one of them with some probability, or could it be a monstrous mixture of all of them?
This seems to demonstrate the point from a few slides ago that one of the challenges of this kind of prediction is how data hungry it is. The training data might need a lot of data points with occluded legs and labels that tell the model what the legs should look like when they're obscured.