Draw the model predicted trajectory

[Alex-fishred]

Can I know the position (pixel) of the model predicted object on the image?
my request:
I want to plot the trajectory predicted by the model

I found that the model outputs only Bayesian probabilities

Is there any chance of it happening?

[Cogito2012]

@Alex-fishred Good question!
We actually use the detected objects to perform the accident probabilities. For your request, you may need to find out which object contributes most to the probabilistic decision. However, since we use a static graph in GCN to aggregate the object features (graph edge not learned), the graph edge still cannot reveal the learned contribution of each node/object.

As a simple way to work around it, you may compute the gradient norm of the object feature (or GCN node feature) as the importance score. The position of the object with the highest score should be what you want :)

[Alex-fishred]

you mean
gradient norm
Equivalent to showing where the model is currently focused?
For example, I have targets that are 5 meters and 10 meters away from the camera
Will the gradient norm obtained at 5 meters be larger than that obtained at 10 meters?

if i'm right
Can you tell me how to implement compute the gradient norm of the object feature (or GCN node feature)