zhixiangwei/ReinPlusPlus

Rein++

Checkpoints for Rein++: Efficient Generalization and Adaptation for Semantic Segmentation With Vision Foundation Models (IEEE TPAMI 2026).

Paper · Inference and evaluation code · Validation report

First checkpoint release

We release the first batch of Rein++ checkpoints and inference and evaluation code for reproducing our paper results. This release includes five Rein-G models based on DINOv3-L and DINOv3-7B, and one Rein-A model, covering experiments in Tables II and VII. All six models have been evaluated on their corresponding validation sets. The remaining checkpoints and training code will be released progressively.

Models

Scores are mIoU (%). Paper results and results measured with the release are listed separately.

Model ID	Method / backbone	Paper table	Paper	Re-evaluated	Download
rdg_usg2cbm_dinov3_rein	Rein-G / DINOv3-L	II	78.9 / 63.4 / 74.8	78.91 / 63.44 / 74.78	config · weights
rdg_usg2cbm_dinov3-7b_rein	Rein-G / DINOv3-7B	II	80.0 / 66.3 / 75.9	80.02 / 66.32 / 75.91	config · weights
rdg_c2abm_dinov3_rein	Rein-G / DINOv3-L	II	72.5 / 65.0 / 77.4	72.52 / 65.04 / 77.41	config · weights
rdg_c2abm_dinov3-7b_rein	Rein-G / DINOv3-7B	II	75.9 / 68.0 / 77.5	75.92 / 67.97 / 77.54	config · weights
rdg_g2cbm_dinov3-7b_rein	Rein-G / DINOv3-7B	II	73.9 / 65.5 / 72.6	73.89 / 65.49 / 72.58	config · weights
rda_g2c_dinov3_rein_ta	Rein-A / DINOv3-L	VII	79.0	79.03	config · weights

Metric order and settings:

• usg2cbm: UrbanSyn + SYNTHIA + GTAV → Cityscapes / BDD100K / Mapillary.
• c2abm: Cityscapes → ACDC / BDD100K / Mapillary.
• g2cbm: GTAV → Cityscapes / BDD100K / Mapillary.
• g2c: GTAV → Cityscapes (domain adaptation).

All scores use validation sets. ACDC is the arithmetic mean of its night, snow, fog, and rain mIoU values. Table II's DINOv3-L GTAV-only model is planned for a later release.

Each model folder contains model.pth, config.py, and manifest.json. The checkpoint contains the Rein adapter parameters and segmentation head from the student branch used for evaluation. Load the matching DINOv3 backbone first, then these parameters. The manifests record complete SHA-256 checksums and the required backbone version; model-index.json collects the model metadata and results.

Download

pip install -U huggingface_hub
hf download zhixiangwei/ReinPlusPlus --local-dir /absolute/path/ReinPlusPlus

To download one model:

hf download zhixiangwei/ReinPlusPlus --include "rdg_usg2cbm_dinov3_rein/*" \
  --local-dir /absolute/path/ReinPlusPlus

For a full repository download, run sha256sum -c SHA256SUMS inside the downloaded directory to check the release files.

Base models

Obtain the original PyTorch LVD-1689M pretrained checkpoints from DINOv3's official repository:

Backbone	File	SHA-256
DINOv3-L	dinov3_vitl16_pretrain_lvd1689m-8aa4cbdd.pth	8aa4cbddda325040fc78db2c272754af6ebe8ff2c55f6ec4f1964d8890f66035
DINOv3-7B	dinov3_vit7b16_pretrain_lvd1689m-a955f4ea.pth	a955f4ea3bec4fcd666bf363630da4386383069b482c8a927e17a3e1154965b7

Use the original PyTorch files with these configurations. Base checkpoints are downloaded separately under Meta's access and license terms.

Inference and evaluation

Inference and evaluation code is available in projects/rein++. Follow the subproject README for installation and data preparation.

The validation environment uses Linux, Python 3.10, PyTorch 2.1.0 with CUDA 11.8, MMCV 2.1.0, MMEngine 0.10.7, MMSegmentation 1.2.2, and MMDetection 3.3.0.

From the Rein++ subproject directory, replace MODEL_ID with an entry in the table:

python tools/infer.py /absolute/path/ReinPlusPlus/MODEL_ID/config.py \
  /absolute/path/ReinPlusPlus/MODEL_ID/model.pth /absolute/path/image.png \
  --backbone /absolute/path/dinov3_pretrain.pth \
  --output-dir /absolute/path/predictions

python tools/test.py /absolute/path/ReinPlusPlus/MODEL_ID/config.py \
  /absolute/path/ReinPlusPlus/MODEL_ID/model.pth \
  --backbone /absolute/path/dinov3_pretrain.pth \
  --data-root /absolute/path/data --work-dir /absolute/path/evaluation

Evaluation also accepts --cfg-options key=value.

Labels and datasets

Predictions use Cityscapes train IDs 0–18, in this order: road, sidewalk, building, wall, fence, pole, traffic light, traffic sign, vegetation, terrain, sky, person, rider, car, truck, bus, train, motorcycle, bicycle. Ground-truth ID 255 is ignored.

The evaluations use Cityscapes validation (500 images), BDD100K validation (1,000), Mapillary Vistas v1.2 validation (2,000), and ACDC validation (406 across four conditions), as selected by each configuration. Mapillary labels use the experiment's mapping to the 19 Cityscapes classes. The code release provides data preparation instructions and label conversion tools. ACDC results here refer to validation, not the official test server.

Training

Training code will be released progressively.

License

The Rein adapter and segmentation-head weights are released under GPL-3.0. DINOv3 base models retain their original license. DINOv3-derived code and OpenMMLab components retain their respective licenses, documented in the code repository.

Citation

@article{Wei2026ReinPlusPlus,
  author = {Wei, Zhixiang and Ma, Xiaoxiao and Yan, Ruishen and Tu, Tao and Chen, Huaian and Zheng, Jinjin and Jin, Yi and Chen, Enhong},
  title = {Rein++: Efficient Generalization and Adaptation for Semantic Segmentation With Vision Foundation Models},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  year = {2026},
  volume = {48},
  number = {8},
  doi = {10.1109/TPAMI.2026.3681631}
}