LibrePPLiteSeg — third-party attribution
========================================

Code
----

The PP-LiteSeg architecture (STDC backbone, SPPM context module, UAFM decoder,
segmentation and auxiliary heads), the Dice + cross-entropy + edge compound
loss, the preprocessing convention, and the Cityscapes training recipe in this
directory are adapted from:

  Deci-AI/super-gradients, commit 63de22c404d5740f34f7706c302b37fce3c8fe5d
  https://github.com/Deci-AI/super-gradients/tree/63de22c404d5740f34f7706c302b37fce3c8fe5d
  License: Apache License 2.0
  https://github.com/Deci-AI/super-gradients/blob/63de22c404d5740f34f7706c302b37fce3c8fe5d/LICENSE.md

  Adapted files:
    src/super_gradients/training/models/segmentation_models/ppliteseg.py
    src/super_gradients/training/models/segmentation_models/stdc.py
    src/super_gradients/training/models/segmentation_models/common.py
    src/super_gradients/training/models/segmentation_models/context_modules.py
    src/super_gradients/modules/conv_bn_act_block.py
    src/super_gradients/training/losses/dice_ce_edge_loss.py
    src/super_gradients/training/losses/dice_loss.py
    src/super_gradients/training/losses/mask_loss.py
    src/super_gradients/training/losses/structure_loss.py
    src/super_gradients/training/utils/segmentation_utils.py
    src/super_gradients/recipes/cityscapes_pplite_seg50.yaml
    src/super_gradients/recipes/cityscapes_pplite_seg75.yaml
    src/super_gradients/recipes/dataset_params/cityscapes_stdc_seg50_dataset_params.yaml
    src/super_gradients/recipes/dataset_params/cityscapes_ppliteseg_seg75_dataset_params.yaml
    src/super_gradients/recipes/training_hyperparams/cityscapes_default_train_params.yaml

STDC backbone lineage. The adapted SuperGradients ``stdc.py`` states that it is
based on the original STDC-Seg implementation, so that notice travels with the
STDC-derived code here:

  MichaelFan01/STDC-Seg, commit 59ff37fbd693b99972c76fcefe97caa14aeb619f
  https://github.com/MichaelFan01/STDC-Seg/tree/59ff37fbd693b99972c76fcefe97caa14aeb619f
  License: MIT
  https://github.com/MichaelFan01/STDC-Seg/blob/59ff37fbd693b99972c76fcefe97caa14aeb619f/LICENSE

Architecture behavior was cross-checked against the original PaddlePaddle
implementation. No PaddleSeg code was copied:

  PaddlePaddle/PaddleSeg, commit 3c4db66de1d9d59d0628ed87590b6308a2f4aa2a
  https://github.com/PaddlePaddle/PaddleSeg/tree/3c4db66de1d9d59d0628ed87590b6308a2f4aa2a
  License: Apache License 2.0

Paper: "PP-LiteSeg: A Superior Real-Time Semantic Segmentation Model",
Peng et al., arXiv:2204.02681.

The LibreYOLO API surface, checkpoint integration, rectangular dataset support,
exporter integration, tests, and documentation are original work for this
repository and are MIT-licensed like the rest of LibreYOLO.

Weights (a separate surface — read this before redistributing)
--------------------------------------------------------------

The released LibrePPLiteSeg{t50,b50,t75,b75}-sem.pt checkpoints are converted
from the SuperGradients Cityscapes-trained artifacts. The Cityscapes license
(https://www.cityscapes-dataset.com/license/) permits distributing abstract
derivative models from which the dataset cannot be recovered, and restricts the
dataset and its derivatives to NON-COMMERCIAL use. Those pretrained checkpoints
are therefore non-commercial, regardless of the permissive code license above.

This does not make LibreYOLO, or the PP-LiteSeg architecture, non-commercial.
Weights trained from scratch on your own data carry none of that restriction; a
fine-tune initialized from a released Cityscapes checkpoint inherits it.

Source artifacts and their SHA-256 digests, verified before conversion:

  t50  https://d2gjn4b69gu75n.cloudfront.net/models/pp_lite_t_seg50_cityscapes.pth
       ae9ad0cae645ebdfb8de661e8dbf1c33e08c0d90997954429b467e70e9ca4194
  b50  https://d2gjn4b69gu75n.cloudfront.net/models/pp_lite_b_seg50_cityscapes.pth
       f7a6769dd37290ee1145c5f1aa2a669b5421591d3afaf8629910614405fe7122
  t75  https://d2gjn4b69gu75n.cloudfront.net/models/pp_lite_t_seg75_cityscapes.pth
       1fdd809572a1b3168727ed0dea32da287c9917ed0ebbfdf8ecab87a3116733f6
  b75  https://d2gjn4b69gu75n.cloudfront.net/models/pp_lite_b_seg75_cityscapes.pth
       383bfb69b3ffb643224dbf76047acacd9b48a1564958a20f1aca69c5453dca3e

Cityscapes itself is user-supplied. LibreYOLO does not download it, accept its
terms on your behalf, or bundle any of its images or labels.
