Version 0.20.1

• Bumped up dependency versions.
• Refactored code and addressed some of the new Cargo Clippy feedback.

Version 0.20.0

Significant accuracy and quality improvements:

• Adjusted cross-correlation parameters to extract more data, and rely on the cross-check filter to remove errors and noise.
• Using matches from fundamental matrix to estimate image pose (improve performance and rely on more solid data than cross-correlation)
• Use a depth buffer to detect polygon obstruction and output rendered images (improves accuracy and significantly boosts performance)
• When merging track, average tracks instead of consuming them; by checking that all tracks follow the same direction, filter outliers
• Removed minimal angle filter (it's not as robust as it seemed originally)
• Fixed focal distance formula (it's supposed to use diagonals, not width)
• Adjust some parameters based on image size, to improve handling of larger images
• Removed requirements for points to be far apart in P3P
• Use point projections to build mesh (instead of original image) - as P3P and bundle adjustment can move cameras

Other changes:

• Better logging
• Option to limit number of points (avoid generating a mesh that doesn't fit into memory)
• Removed point normals, as they just used extra memory and caused artefacts in the macOS preview

Version 0.19.2

Small improvement release:

• Bumped up dependency versions
• Switched to Github's ARM runners for macOS, recompiled Metal shaders
• Refactored error handling code
• Addressed Clippy comments

Version 0.19.1

• Improved support for low-power GPUs (fixed regression from 0.19.0)
• Fixed the help text
• Updated to Rust 1.77
• Updated dependency versions, including nalgebra to fix accuracy in some use cases

Version 0.19.0

Refactored code and reduced number of external dependencies:

• Adjusted a few parameters, ensure they scale correctly with image size
• Using a custom Grid type instead of nalgebra DMatrix
  • Avoid weird conversion between row/column and x/y coordinates
  • Allow for custom optimizations in the future (if needed)
• Replaced wgpu with a lower-level GPU implementation
  • Reduced binary size and number of dependencies
  • Use GPU APIs in the most optimal way possible (e.g. copy data directly to GPU if it shares memory with the host)
  • Precompile & validate shaders during the build, instead of during runtime
  • Keep GPU through the entire process - skip re-loading the Vulkan library
• Reduced image correlation memory usage by 50%
• Tweaked correlation parameters for perspective projection - less data, but higher quality
• Fixed small bugs in GPU shaders implementation
• Fixed vertex color output for OBJ files

Some of the changes might have fixed historical, difficult-to-find bugs and improved overall robustness.

Version 0.18.0

Improvements in accuracy and general output results:

• Optimize the fundamental matrix and use more matching points for validating it
• Use a peak/noise filter from the Photo Tourism paper; while it reduces noise, it will discard points that are too far away
• Match all images with each other, instead of just adjacent ones
• Choose the best initial image pair
• Fixed polygon obstruction detection when constructing a mesh from more than one image
• When determining camera pose, prefer points that are far apart
• Improve image output
  • Add polygon point normals
  • Deduplicate polygons that are visible on multiple images and connect the same points
  • Output texture coordinates for all images, provide a bit more attributes in the Alias|Wavefront OBJ material files

Version 0.17.0

Significant improvements in reconstructing images with perspective projection.

• Switched to 7-point algorithm for detecting the Fundamental Matrix
• Optimize Fundamental Matrix using Levenberg-Marquardt least squares
• Match keypoints using ORB descriptors instead of using cross-correlation
• Use focal length from EXIF data to create an accurate projection matrix
• Use analytical formula for rotation matrix (and switch to Rodrigues' formula)
• Several bugfixes in sparse bundle adjustment significantly improve the end result (SBA finally works now!)
• Sparse bundle adjustment runs several times faster, without using additional RAM
• Detect and discard backprojected points
• Improved outlier filter
• Create interpolated meshes for SFM (more than 2 images)
• Windows version no longer requires the Visual C++ Runtime
• Reduced the binary size

Version 0.16.0

• Improved reconstruction of photos (perspective projection)
• Added experimental support for structure-from-motion (reconstruct a 3D object from multiple views)
  • Results can be noisy and sometimes fail
  • Interpolation only works when 2 or less images are reconstructed

Version 0.15.1

Bumped up dependency versions.

Version 0.15.0

Significant improvements in performance and accuracy:

• A new peak filter now eliminates noise more effectively, without discarding good data points
• A few adjustments in the algorithm increase speed (up to 2x-3x) without affecting accuracy

Finally, perspective projection (for regular photos) now works - with real projection matrices instead of a "disparity" map.