gtsfm
gtsfm copied to clipboard
borglab
End-to-end SFM framework based on GTSAM
- Clarify in the HTML report what determines a "valid" image pair - Have 2 separate thresholds for R and U success? Put in place right before the results against...
We have runners for specific datasets, but there is no simple way right now to just run GTSFM on a folder of images. We can solve this using the following:...
missing access to (1) image filenames (2) unique 2d point indices (3) camera id vs. image id in colmap-format dump/ write
We have one method for shared intrinsics currently (via separated intrinsics and poses): https://github.com/borglab/gtsfm/pull/65 This method slows down BA, presumably due to increased fill-in. We propose to have a second...
- Consider if we need more RANSAC iterations - Consider if we need more random 1dsfm projection directions - Analyze how much a random seed on the items increases determinism....
Currently, the inlier ratio with respect to GT model is greater than that with respect to estimated model for USAC. This is perhaps expected, as the GT should explain the...
- Compare how Theia'a method of filtering by simulated pairwise (from global) vs. actual pairwise measurements helps - Add to HTML plot the front-end inputs after each outlier rejection step...
Make pixel thresholds a function of the image size, to avoid treating 480p and 1296p images differently. A good reference is http://phototour.cs.washington.edu/ModelingTheWorld_ijcv07.pdf: We set the RANSAC outlier threshold to be...
So currently we have k1 and k2 hard coded in different I/O reading files, and as default for EXIF: https://github.com/borglab/gtsfm/blob/master/gtsfm/common/image.py#L68 https://github.com/borglab/gtsfm/blob/master/gtsfm/utils/io.py#L189 They are also hard-coded to zero when resizing intrinsics....
npm start failing due to NaN values in 1dsfm_outlier_angular_errors_deg, averaging_metrics.json
