Add bal dataloader to test SBT

[hxu296]

This PR adds a BAL dataset loader at pypose/examples/sparse/bal_loader.py. fix: #235

This change includes 4 components:

1. BAL data loader written in TorchData’s functional style: bal_loader.py
2. BAL specific util functions for calculating reprojerr and visualizing loss: bal_utils.py
3. Dense bundle adjustment solver using Adam optimizer: ba_dense.py
4. Google Colab example, so anyone can try it painlessly: ba_example.ipynb (Demo)

[hxu296]

As discussed with Zitong, I will add a PyTorch SGD based BA algorithm to test the dataloader and also serve as a simple baseline for sparse LM based BA.

Also thank Zitong for going through the SBT implementations and sparse Jacobian plans with me, which helps me to catch up to speed :)

[huyaoyu]

Only reviewed the bal_loader.py. Need some more time on ba_dense.py.

[hxu296]

Thanks for your review @huyaoyu
ba_dense.py is still WIP as for now. I will ping you when it's ready for review, thanks!

[hxu296]

The CI failure looks weird, as commit 535e907 is scoped within examples/sparse...

[hxu296]

Addressed all comments regarding bal_loader.py. Added a new file bal_utils.py for a custom reprojerr function that considers radial distortions k1 and k2.

These two files are ready for review:

1. bal_loader.py
2. bal_utils.py

Meanwhile, I will keep working towards a bal_dense.py and a jupyter notebook example that is colab ready. Cheers.

[hxu296]

Hi, the following two files are ready for review, and nothing else is on the workbench.

1. ba_dense.py
2. ba_example.ipynb

You can try the jupyter notebook example through google colab by clicking on the "Open in Colab" tab in the README.

Here is a visualization of loss before and after optimization:

Here is a visualization of the loss history:

Cheers.

[huyaoyu]

Some simple comments.

Need a quick call for discussion.

Thank you!

[huyaoyu]

Let's have a quick meeting to discuss how to do proper comments to this function.

I have a hard time understanding the magic dimension numbers used in this function. I think future users might try to take a look at how projection errors are implemented and they might run into the same problem.

[hxu296]

After a discussion, Yaoyu and I reached the conclusion that we should integrate the image undistortion function into pypose's geometry module, and then simply call pypose native functions here. This will avoid exposing the complexity of raw dimension manipulations to the users and also strengthen pypose's API library.

A refactor is on the way.

[hxu296]

Hello! Thanks for your review. I am up to a call after the monthly developer meeting through Slack or Zoom. My email is [email protected].

Update: we had a nice in person discussion