LibreDOMEDETR includes code ported from Dome-DETR:

Source: https://github.com/RicePasteM/Dome-DETR
Commit: 2dde3bc1946a3e9fad9abd0612b59fc39bd6b861
License: Apache License 2.0.
Copyright (c) 2025 The Dome-DETR Authors. All Rights Reserved.

The upstream LICENSE file is the standard Apache-2.0 text but leaves the
copyright line as the "[yyyy] [name of copyright owner]" placeholder, so there
is no named holder in the file itself. The attribution above is taken from the
upstream source headers.

The upstream implementation is itself modified from D-FINE:

Source: https://github.com/Peterande/D-FINE
License: Apache License 2.0.
Copyright (c) 2024 The D-FINE Authors. All Rights Reserved.

LibreYOLO ports the three modules that distinguish Dome-DETR from D-FINE: DeFE
(the density-focal extractor, src/zoo/dome/defe.py), MWAS (the masked window
attention processor, src/zoo/dome/get_roi_features.py), and PAQI together with
its density-adaptive NMS (src/zoo/dome/dome_decoder.py and
src/zoo/dome/dynamic_nms.py). The static-shape MWAS formulation follows the
same repository's "onnx-export" branch.

Everything Dome-DETR shares unchanged with D-FINE (the HGNetv2 backbone, the
CSP/RepNCSPELAN4 encoder blocks, the FDR decoder stack, and the pure-PyTorch
multi-scale deformable attention) is imported from libreyolo/models/dfine/
rather than re-vendored here, and is covered by that family's attribution.

The training path is ported too: the Dome-DETR criterion (loss.py, a subclass
of D-FINE's), the DeFE density and count supervision together with the
ground-truth density-map construction it is trained against (encoder.py), and
the contrastive-denoising adaptation for a variable per-image query count
(denoising.py). Each of those files carries the attribution above in its own
header.

Not included: upstream's data pipeline, which LibreYOLO replaces with its own;
the deformable-encoder branch (no released config enables it); and the
prebuilt MSDeformAttn CUDA extension under src/zoo/dome/ops/. The Hungarian
matcher is inherited from libreyolo/models/dfine/ rather than re-vendored here,
and is covered by that family's attribution.

Model weights are not redistributed by LibreYOLO. See
weights/LICENSE_NOTICE.txt for why.
