Travis Driver

@BangwenHe This fixed it for me, thank you! Specifically, I used pytorch 0.4.1 and python 3.7.

Yes, you're right. I was under the impression that the MVS experiments used the estimated poses as opposed to GT. PixSfM also demonstrated impressive pose estimation results on ETH3D, but...

Also, are you sure inlier weighting was turned on it master when you compared? I vaguely remember @akshay-krishnan accidently turning it off in master.

# Shonan weighted by # of inlier matches (IW-Shonan) Huge improvements across the board. **The CI failed for Place of Fine Arts due to memory issues during the Shonan optimization...

@dellaert Is there any other way to enable the Huber loss for Shonan other than the `setUseHuber` parameter? From what I can tell, the current API doesn't allow us to...

# IW-Shonan w/ Huber Loss Huber is still not effective.  [visual_comparison_dashboard.zip](https://github.com/borglab/gtsfm/files/13735517/visual_comparison_dashboard.zip)

Some unit tests failing to do `FAILED tests/test_gtsfm_runner_olssonloader.py::test_gtsfm_runner_olssonloader[dataset_path0] - hydra.errors.MissingConfigException: Cannot find primary config 'sift_front_end.yaml'. Check that it's in your config search path.`

It looks like you deleted a couple of the front-end configs. Did you mean to do that?

### `robust_estimation_type = USAC_ACCURATE`, `estimation_threshold_px = 0.5`  [visual_comparison_dashboard.html.zip](https://github.com/borglab/gtsfm/files/13350550/visual_comparison_dashboard.html.zip)

> I think we can improve the design a bit, staying closer to gtsfm style. > > First, I think descriptors can remain separate from keypoints. Second, are the input...