Submit Results — Industrial3D

Option A (recommended)

Submit via Pull Request

The canonical way. Your submission is transparent, versioned, and linked to your GitHub identity.

1

Fork the repository

Fork github.com/PointCloudYC/Industrial3D to your GitHub account.

2

Edit `website/data/leaderboard.json`

Append a new entry object to the "entries" array. See the required JSON schema below.

3

Open a Pull Request

Open a PR against the main branch of the upstream repo. Title it:

[Submission] MethodName — XX.XX% mIoU

In the PR body, include:

Link to paper or preprint (arXiv preferred)
Link to code repository (required for verification)
Brief description of the method
Confirmation that you followed the standard evaluation protocol

4

Review & merge

The Industrial3D team will verify your results (code + reproducibility check) and merge the PR. The leaderboard updates automatically on the next deployment.

Open a Pull Request →

Option B

Submit via GitHub Issue

For those not comfortable editing JSON. Open an issue with the Benchmark Submission template and we will add your results.

1

Open a Benchmark Submission issue

Use the prefilled template at:
New Issue → Benchmark Submission

2

Fill in all required fields

Method name, paradigm, label budget, mIoU, paper link, code link, and evaluation protocol confirmation.

3

We transcribe & add

The team reviews and manually adds your entry to the leaderboard JSON within ~7 days.

Open Issue Template →

Format

Required JSON Entry Schema

Add one object to the "entries" array in website/data/leaderboard.json. Required fields are marked with *.

leaderboard.json — entry object

{
  "id":        <next integer>,          // auto-increment from last entry
  "method":    "YourMethodName",         // * short, recognisable name
  "paradigm":  "Fully-Supervised",       // * one of the four paradigms below
  "labels":    "100%",                   // * label budget used ("0%" for zero-shot)
  "miou":      52.34,                    // * mIoU as a decimal number (not string)
  "oa":        null,                     // overall accuracy (null if not reported)
  "year":      2026,                     // * paper or preprint year
  "paper_url": "https://arxiv.org/...", // * link to paper/preprint
  "code_url":  "https://github.com/...", // * public code repo (required for verify)
  "submitter": "Your Name",             // * your name or team name
  "verified":  false,                    // set to false; authors flip to true
  "per_class": {                         // optional but strongly encouraged
    "Duct": 89.1, "Elbow": 22.3, "Flange": 30.1,
    "IBeam": 95.2, "Pipe": 72.5, "Pump": 5.1,
    "RBeam": 93.0, "Reducer": 0.0, "Strainer": 0.0,
    "Tank": 88.7, "Tee": 4.2, "Valve": 33.6
  },
  "notes": "Optional short note about the method or evaluation."
}

Valid paradigm values

Fully-Supervised Weakly-Supervised Unsupervised Foundation Model

If your method doesn't fit cleanly, use the closest paradigm and clarify in the PR description or notes field.

Protocol

Evaluation Protocol Details

Test Set

Held-out areas 6 (93m PSU / 99.2 VSPA-1) and 12 (Sludge Press House). 84.9M points, 13.9% of the dataset. Area 9 (15.1M points) is reserved for validation only — do not report test-set numbers tuned on Area 9.

Standard Evaluation

Vote-based smooth testing (test_smooth = 0.95) following the RandLA-Net / SQN convention. Run multiple passes over the test set and accumulate per-point logits before final argmax. Report the 12-class mIoU on the full test set.

Label Budget (Weakly-Supervised)

Specify the exact fraction of training-set points used as supervision. We report results for 0.01% and 0.1% budgets; other budgets are welcome. Pseudo-label variants should be described in the notes field.

Foundation Model / Zero-Shot

Report the setting (e.g., zero-shot, few-shot N%, oracle matching) clearly. If using class name prompts or GT mask proposals, specify this in the notes field. Hungarian matching results should be accompanied by the matching protocol.

Citation

Cite Industrial3D

If you use Industrial3D in your research, please cite the dataset paper. We also ask that you cite the related work below if you compare against those methods.

Primary — Industrial3D Dataset Paper

@article{yin2026industrial3d,
  title={Industrial3D: A Terrestrial LiDAR Point Cloud Dataset and
         Cross-Paradigm Benchmark for Industrial Infrastructure},
  author={Yin, Chao and Yue, Hongzhe and Han, Qing and Hu, Difeng and
          Liang, Zhenyu and Lin, Fangzhou and Sun, Bing and Wang, Boyu and
          Li, Mingkai and Yao, Wei and Cheng, Jack C.P.},
  journal={arXiv preprint arXiv:2603.28660},
  year={2026}
}

Related — Companion Methods Paper (Dual Crisis Solutions)

@article{Yin2026arXiv,
  title={Resolving Primitive-Sharing Ambiguity in Long-Tailed Industrial
         Point Cloud Segmentation via Spatial Context Constraints},
  author={Yin, Chao and Han, Qing and Hou, Zhiwei and Liu, Yue and
          Dai, Anjin and Hu, Hongda and Yang, Ji and Yao, Wei},
  journal={arXiv preprint arXiv:2601.19128},
  eprint={2601.19128},
  year={2026}
}

Related — ResPointNet++ / PSNet5 (Prior Supervised Baseline)

@article{yin2021,
  title={Automated semantic segmentation of industrial point clouds
         using ResPointNet++},
  author={Yin, Chao and Wang, Boyu and Gan, Vincent JL and Wang, Mi
          and Cheng, Jack CP},
  journal={Automation in Construction},
  volume={130},
  pages={103874},
  year={2021},
  publisher={Elsevier},
  doi={10.1016/j.autcon.2021.103874}
}

Related — SQN (Prior Weakly-Supervised Baseline)

@article{yin2023,
  title={Label-efficient semantic segmentation of large-scale industrial
         point clouds using weakly supervised learning},
  author={Yin, Chao and Yang, Bo and Cheng, Jack CP and Gan, Vincent JL
          and Wang, Boyu and Yang, Ji},
  journal={Automation in Construction},
  volume={148},
  pages={104757},
  year={2023},
  issn={0926-5805},
  doi={10.1016/j.autcon.2023.104757}
}

Submit Your Results