LibrePPLiteSegb50-sem

NON-COMMERCIAL WEIGHTS. These weights are trained on Cityscapes. The Cityscapes license permits distributing abstract derivative models from which the dataset cannot be recovered, and restricts the dataset and its derivatives to non-commercial use. The restriction applies to this checkpoint, not to LibreYOLO's MIT code or to the PP-LiteSeg architecture. Weights you train from scratch on your own data carry none of it; a fine-tune started from this checkpoint inherits it. Read the restriction before you download.

PP-LiteSeg b50 (STDC2 backbone, native 512x1024 canvas), a real-time semantic segmentation model for Cityscapes' 19 classes, repackaged for LibreYOLO.

from libreyolo import LibreYOLO

model = LibreYOLO("LibrePPLiteSegb50-sem.pt")
result = model.predict("street.jpg")[0]
mask = result.semantic_mask.data          # (H, W) class IDs on the original canvas

This is a genuinely rectangular model: it runs at 512x1024 (height x width), not a square canvas. Published Cityscapes validation mIoU for the source checkpoint is 76.48. LibreYOLO has not re-measured that number; it is quoted from the source release, and exact raw-logit parity does not by itself reproduce an end-to-end dataset evaluation.

Source

Derived from Deci-AI/super-gradients at commit 63de22c404d5740f34f7706c302b37fce3c8fe5d. Copyright (c) 2021-2024 Deci AI. Licensed under the Apache License 2.0.

Source artifact: pp_lite_b_seg50_cityscapes.pth, SHA-256 f7a6769dd37290ee1145c5f1aa2a669b5421591d3afaf8629910614405fe7122 (verified before conversion).

The STDC backbone lineage comes from MichaelFan01/STDC-Seg at commit 59ff37fbd693b99972c76fcefe97caa14aeb619f, MIT. Copyright (c) 2021 Mingyuan Fan. The architecture was cross-checked against PaddlePaddle/PaddleSeg at commit 3c4db66de1d9d59d0628ed87590b6308a2f4aa2a, Apache-2.0; no PaddleSeg code was copied.

Paper: PP-LiteSeg: A Superior Real-Time Semantic Segmentation Model.

Modifications

State-dict key remapping only: the upstream net payload with exactly one module. DDP prefix stripped, wrapped in LibreYOLO v1.0 checkpoint metadata. Learned parameters are unchanged, and the three training auxiliary heads are retained so the checkpoint stays trainable. The port reproduces the pinned upstream exactly (max_abs_diff == 0.0 on the main logits). See weights/convert_ppliteseg_weights.py and weights/parity_ppliteseg.py in the LibreYOLO source repository.

License

The code is Apache-2.0 (super-gradients) with MIT STDC lineage; both license texts are in LICENSE, and attribution is in NOTICE.

The weights in this repository are non-commercial under the Cityscapes dataset terms linked above. Downstream users are responsible for complying with them.