- Supports python[3.9-3.13]
- Fixed issue #22 by releasing Numpy lock.
- Fixed issue #19, by referencing the indicies from the score_mask array
- Dropped support to python 3.7
- Better handling of the dev dependencies in the setup file
- Compatible with python 3.10
- Supports Numpy 1.24
- Fixed issue #12, by masking scores as well as boxes.
- Added torch and torchvision as proper dev dependencies
- Fixed Pillow version (dev dep) to 9.3.0 in dev dependencies because 9.4.0 does not compile on my mbp (see python-pillow/Pillow#6862)
- Removed deprecated arguments:
cutoff_distanceandtree. Removed associated tests. - Added sanity check to ensure
leaf_sizeis strictly positive.
- Edge case where all box scores are zero (or all below threshold) is now handled (threw uggly error before)
- Added multiclass support for NMS
- Use the underlying RNode class in nms now, which speeds up compilation (no need to compile RTree anymore)
- Discarded BallTree and KDTree which are now replaced by a RTree: as fast and hyperparameter-free
->
lsnms.nmsis now twice faster to compile (only one tree to compile) cutoff_distanceandtreeare now deprecated. Warnings are issued when those are specified.- cleared the tests structure
- added types conversion and sanity checks for data shape, etc
- cached all the jitted function which could (the non recursive ones) -> saves 2 seconds of compilation time at first use.
->
lsnms.nmsis now hyperparameter free, runs just as fast as before, and almost three times faster to compile at first use.
- Fixed typo in BallTree (missing dimensionality argument, not used in algorithm but for sanity check in query)
- Updated Numba to 0.54.1
- First version
- Both BallTree and KDTree are implemented for the sake of exhaustivity
- A cutoff distance needs to be specified to discard boxes to distant one from the other
- Compilation time at first use is quite long: 13 seconds and functions can not be precompiled due to recursivity.