Closed hawesie closed 9 years ago
@hawesie: @ferdianjovan didn't want to do what you sugest, as the msg might become too large given a stationary person. Do you have an opinion on that?
Yup, that is what I want to avoid. If we include information from human trajectory classifier, we might be able to drop stationary persons from the message.
I meant internally to the recognition node. But if it's a problem there then ignore this.
I can hold onto the qsrs for the previous uuid inside the service (if you agree) and check whether it's the same uuid as the current uuid. I should be able to calculate the qsrs for the new batch, and then add them onto the end of the matrix (obj-pairs x qsr per time point). Assuming services can remember things between calls...
@ferdianjovan Do the mini-batches of trajectories you publish ever overlap? Because I receive the first mini-batch trajectory as: poses[0] -> poses[n], then the second mini-batch poses[0] -> poses[m]. Can I assume poses[0] in the second mini-batch is timepoint [n+1] ??
Yes, you can
@hawesie, @ferdianjovan This is done. Clone my repo for updated code.
@PDuckworth please can you create a PR to add the updated code to the copy of your repo in the strands-project account. We should start treating that as the master repo, with people forking from that for development.
@hawesie Done.
Let me know if it doesn't work. It should be a stable version.
@hawesie can we close this? They are being stitched together after the QSRs are being generated. The Episodes are updated on mongodb if they belong to the same UUID.
The mini-batch trajectories come in non-overlapping chunks. In the recognition node these can be stitched together to make longer trajectories.
(as an aside, maybe @ferdianjovan can just make the mini-batch trajectories extend as more data comes in)