V1.6.13 - Fix

Fixes possible side effects when changing probas to save at predict time.

What's Changed

• Add bds3 licence by @CharlesGaydon in #9
• Setup up GitHub Pages hosted documentation by @CharlesGaydon in #10

Full Changelog: V1.6.11...V1.6.13

V1.6.11

Output LAS format is fixed to point format 8.

Refer to https://github.com/IGNF/lidar-deep-segmentation/releases/tag/V1.6.3 for trained model and config assets.

What's Changed

• HOTFIX : use pdal instead of laspy for interpolation by @CharlesGaydon in #8

Full Changelog: V1.6.3...V1.6.10

V1.6.10

Code is now robust to extra bytes that are created during lidar acquisition or by TerraScan.

Refer to https://github.com/IGNF/lidar-deep-segmentation/releases/tag/V1.6.3 for trained model and config assets.

What's Changed

• HOTFIX : use pdal instead of laspy for interpolation by @CharlesGaydon in #8

Full Changelog: V1.6.3...V1.6.10

V1.6.3

This makes the 50% overlap between successive receptive fields the default. Inference takes longer (times (1/0.5)² = 4), but prediction are more homogeneous: there are fewer artifacts and less noise.
A new building model is attached that was trained using GridSampling, with equivalent performances (86% IoU), for the sake of coherence with data preparation at predict time.

Full Changelog: V1.6.2...V1.6.3

V1.6.2

Changes :
Test and Predict now share a common logic based on KNN interpolation of all available inference results, in a way that is agnostic to how they were obtained and if there are unique or not. Testing a model enables a full evaluation of the IoU without approximation, as labels for all points are used instead of those of subsampled points. Therefore, various combinations of cloud subsampling, interpolation methods, and subtile overlaps can be compared in a robust way.

Discussion:
We observe that with a previously trained model (same as in V1.0.0), KNN with k=10 points at inference time combined with grid sampling can improve IoU of up to 1.5%pts at almost no cost. This config therefore becomes the new default. Best improvements come from interpolation using KNN (k=10) on overlapping subtitles (up to 5%pts of improvement). This has at a high cost, virtually doubling (if overlap=25m) or quintupling (overlap=40) the number of needed data pass.

Full Changelog: V1.6.0...V1.6.2

V1.0.0

What's Changed

Functional training/prediction ready-to-use repo.

Full Changelog: https://github.com/IGNF/lidar-deep-segmentation/commits/V1.0.0