Official code release for our CVPR 2026 paper, NexusFlow: Unifying Disparate Tasks under Partial Supervision via Invertible Flow Networks.

• 2026.02   NexusFlow is accepted to CVPR 2026. 🎉
• 2025.12   arXiv preprint released: arXiv:2512.06251.

When you train a multi-task BEV perception model and each sample only carries labels for some tasks (e.g. dense mapping labels here, sparse tracking labels there), the partial-supervision signal can drag the shared representation apart. NexusFlow is a lightweight, plug-and-play module that sits alongside the main model: two surrogate networks read the shared BEV feature, squash it with deformable attention, push it through an invertible RealNVP-style coupling layer, and align the two outputs with a simple MSE loss. No architectural change to your main model, no extra labels required.

See materials/readme.md for the architecture and pseudo-code, and materials/NexusFlow.py for a standalone runnable implementation.

Each sample carries labels for only a subset of tasks. Without intervention, the partial-supervision signal pulls the shared BEV representation in incompatible directions; NexusFlow aligns the per-task views of the feature without requiring any additional labels. A higher-resolution version is available as nexusflow_probsetup.pdf.

opensource_code/
├── README.md                          # this file
├── nexusflow_probsetup.png            # problem-setup figure (rasterized)
├── nexusflow_probsetup.pdf            # problem-setup figure (vector)
├── base.py                            # → mmdet/.../detectors/base.py
├── builder.py                         # → mmdet3d/datasets/builder.py
├── dataset_wrappers.py                # → mmdet3d/datasets/dataset_wrappers.py
├── epoch_based_runner.py              # → mmcv/runner/epoch_based_runner.py
├── mmdet_train.py                     # → UniAD/.../uniad/apis/mmdet_train.py
├── nuscenes_e2e_dataset.py            # → UniAD/.../datasets/nuscenes_e2e_dataset.py
├── data_classes.py                    # → nuscenes/eval/detection/data_classes.py
├── evaluate_single_city_eval.py       # → nuscenes/eval/tracking/evaluate.py
└── materials/
    ├── NexusFlow.py                   # standalone NexusFlow module + demo
    ├── readme.md                      # standalone NexusFlow doc
    ├── Baseline_with_map.avi          # qualitative video — baseline
    └── Ours_with_map.avi              # qualitative video — ours

The whole stack is built on top of UniAD. We do not repackage UniAD's environment here — please first follow the upstream UniAD installation guide to set up the conda environment (we use Python 3.9, env name uniad2.0) and clone UniAD to /root/UniAD. Once that works, drop the files in this folder into the paths listed below.

Each file in this folder is a modified copy of an upstream file. Replace the file at the target path with the one here.

The two /root/UniAD/... paths live inside the UniAD repo itself; the mmdet / mmdet3d / nuscenes paths live inside your conda env's site-packages; mmcv is a source-installed clone at /root/mmcv. Adjust the /root/miniconda3/envs/uniad2.0/... prefix to wherever your own environment lives.

• evaluate_single_city_eval.py restricts the nuScenes tracking eval to a single city. Look for the block marked # <<< MOD: filter both pred & gt to only samples in target city (around line 106) and change target_city (default singapore) for a different split.
• data_classes.py mirrors the nuScenes detection eval data classes; it is included because the eval stack imports from it.

pip install torch deformable-attention-pytorch
python materials/NexusFlow.py

This builds the module, runs a forward/backward pass on a random BEV feature, and prints the alignment loss.

If you find this work useful, please cite:

@article{lin2025nexusflow,
  title={NexusFlow: Unifying Disparate Tasks under Partial Supervision via Invertible Flow Networks},
  author={Lin, Fangzhou and Wang, Yuping and Guo, Yuliang and Huang, Zixun and Huang, Xinyu and Zhang, Haichong and Yamada, Kazunori and Tu, Zhengzhong and Ren, Liu and Zhang, Ziming},
  journal={arXiv preprint arXiv:2512.06251},
  year={2025}
}

This repository builds directly on UniAD, mmdetection3d, mmcv, and the nuScenes devkit. We thank the authors for releasing their code.