BC03 3DoF, 1 Sphere, 6D Flow
At test time, we just keep the first 3 dimensions of flow (which are translations) and run the average of it.
Created on May 16|Last edited on May 17
Comment
There are several variants we can try for this:
- One is to just directly output 6D flow by changing the output dimension to 6. This is what I did for the first 5 runs launched in the evening of 05/16/2022. Remember, flow visualizations are only showing translations (though we should show the flow visualizations from the rotations to check that it is the identity).
- (Dave suggests) Another is to enforce more separation by adding separate MLPs for the rotation and translation parts (which may cause further decoupling).
Note that in both cases, if we are doing 3DoF, we are ignoring any rotation at test time, but the rotation is still used as part of the pointwise loss since we need (R,t), so the network should be incentivized to learn the identity rotation.
Compare this with:
- PN++ averaging layer (which has no rotations to worry about, just avearge translations).
- SVD methods which use 3D flow, to see what the difference is.
NOTE: this is after https://github.com/Xingyu-Lin/softagent_rpad/commit/7f30fea29b213c6bbc1dc01b84b4c823e41f7335 which is before Eddie's changes to improve flow extraction (but his was more for SVD if we have to use rotation at test time, I am using the 'mean' action type as it's 3D, so these results should still be valid).
Section 1
Run set
5
Let's try and add some comparisons with prior methods.
6D Flow SVD
5
Avg Layer
7
3D Flow SVD
5
6D Flow
5
Avg Layer
7
3D Flow SVD
0
Add a comment