UrbanTwin Sim2Real LiDAR Challenge: weights of our LUMPI and V2X-Real entries

This repository holds the trained weights and submitted files of our ranked entries (team sadbodh, Sadbodh Sharma) in the UCF UrbanTwin Sim2Real LiDAR Challenge at the IEEE BigData Cup 2026.

Track Final entry Official place Combined score 3D mAP@0.5 Realism
LUMPI lumpi_v62d 3rd of 44 teams 0.5377 0.1885 0.9402
V2X-Real v2x_v35d 3rd of 46 teams 0.8788 0.5825 0.9488

Scores are from the final leaderboards (Codabench competitions 16684 and 16685).

Contents

  • <track>/checkpoints/<model>/: the checkpoint, its config (<model>.yaml) and its train log(s).
  • <track>/submission/submission.zip: the file we uploaded.
  • <track>/submission/anchors/: earlier uploads' predictions.json, which the build scripts take as fixed inputs.
  • lumpi/rescorer/lumpi_rescorer.pkl: the LUMPI Person rescorer.
  • MD5SUMS: the md5 of every file.

LUMPI (lumpi/checkpoints/)

Folder Model Epoch
pp_lumpi_long PointPillars 20
pp_lumpi_d50 PointPillars, density-robust 20
cpv_lumpi_long CenterPoint 15
pp_lumpi_long_all PointPillars, all frames 20
pp_lumpi_br PointPillars BR 20
pp_lumpi_walk PointPillars, walkers 20
pp_lumpi_walk_all PointPillars, walkers, all frames 20
vrcnn_lumpi_fix Voxel R-CNN (L11) 15
vrcnn_lumpi_rs Voxel R-CNN (L11-RS) 15
vrcnn_lumpi_c Voxel R-CNN (L12) 15
vrcnn_lumpi_b_s2 Voxel R-CNN (L11b seed 2) 15
pv_lumpi PV-RCNN++ 15
vrcnn_lumpi_fix_s2 Voxel R-CNN (L11 seed 2) 15
spec_lumpi_person_rs Voxel R-CNN, Person only (L12-Person) 15
vrcnn_lumpi_b Voxel R-CNN (L11b) 15

V2X-Real (v2x/checkpoints/)

Folder Model Epoch
pp_v2x PointPillars 30
cp_v2x CenterPoint 30
pp_v2x_ped PointPillars, riders as persons 30
pp_v2x_all PointPillars, all frames 30
pp_v2x_walk PointPillars, articulated walkers 30
vrcnn_v2x Voxel R-CNN (X6) 20
vrcnn_v2x_b Voxel R-CNN (X6b) 20
vrcnn_v2x_neg Voxel R-CNN, street-furniture negatives 20
fshnet_v2x FSHNet 20
safdnet_v2x SAFDNet 20
dsvt_v2x DSVT-Pillar 20
vmamba_v2x Voxel Mamba 20
pv_v2x PV-RCNN++ 20

How to use

  1. Clone the code repository.
  2. Put each checkpoint at output/pcdet/<model>/default/ckpt/checkpoint_epoch_<N>.pth.
  3. Follow lumpi/README.md or v2x/README.md there.

To check a submitted file byte for byte, run tools/verify_submission.py check --track lumpi|v2x --zip <track>/submission/submission.zip.

Compute: models were trained on an RTX 3090 and on A100 GPUs.

Intended use

These weights are released to reproduce and study our challenge entries; they are research artefacts. They were selected for the two challenge sites and were not tested for any deployment.

Real-data use

The text below is the "Real-data use" section of the reports we sent to the organisers on Nov 15, copied word for word, with LUMPI item 5 added on Nov 18 (it is also in our paper).

LUMPI track

No detector or second stage was trained on real frames or real labels; every model trains on the UT-LUMPI twin. Real data is the public LUMPI Measurement 4 with its labels. No frame in forbidden_frames.txt, and no frame within 20 frames of one, was used for anything except inference on the test frames.

  1. Evaluation and selection. Labelled splits score every change against the gate: val (320 frames below frame 6400) and fresh (320 frames). A third split, NT, has 196 labelled frames: "before" (76 frames, 6401–6476) and "inside" (120 frames between 6765 and 8047). The inside frames are not forbidden, and each is at least 21 frames from a forbidden frame, but they lie within the span of the forbidden test frames. Three changes were chosen on the inside segment:

    • the two-wheeler Person copies, which are in our ranked package lumpi_v62d; we kept them by our own choice, although they failed the val and "before" conditions of the gate;
    • the Van member L12-Van, in lumpi_v61d, which was uploaded after lumpi_v60 and is not our ranked package;
    • the Person refiner L12-Person, in our ranked package lumpi_v62d, which we kept by our own choice after it failed the val condition of the gate.
  2. Box-size statistics. The RS training data rescales synthetic Car, Van, Truck and Bicycle objects to size distributions taken from the labels of 160 real frames (realism_fit). Only the sizes are used; no real point or box enters training.

  3. Constants fitted on val labels. These are the Person size step (+0.10 m height, +0.044 m length, +0.023 m width), the height and length rules, the point-count rules T5 and V5, and the static-Bicycle rule.

  4. Unlabelled real points. The same 160 realism_fit frames, as points only, give:

    • the static-background map P3, which lowers Person and Bicycle scores Γ—0.3;
    • the targets of the realism calibration of our 50 synthetic frames.

    160 further unlabelled frames scored the calibration setting. Two learned point refiners for the realism frames were trained against the same 160 unlabelled frames and read on the other 160; neither passed, so neither is in the final package.

  5. Sensor rig. The walkers and riders added to the training twins are ray-cast through the five-sensor LUMPI rig. Its beam tables come from LUMPI's public sensor metadata, and its five sensor positions were solved from the per-point range values of one real frame (Measurement 4, frame 0, far from any forbidden frame); no labels were used.

Validation, model selection, box-size statistics and realism calibration are on the organisers' list of allowed real-data uses. P3, the constants fitted on val labels and the choices made on the NT inside segment are not on that list as such; we list them here so that they can be judged.

V2X-Real track

No detector or second stage was trained on real frames or real labels; every model trains on the UT-V2X-Real-IC twin. Real V2X-Real data entered our entry in the ways listed below. No frame in forbidden_frames.txt, and no frame within 20 frames of one, was used for anything except inference on the 50 test frames.

  1. Evaluation and selection. Labelled frames score every change against the gate: val_i1 (146 frames of the V2X-Real val split, infrastructure sensor OS1-64) and fresh (328 frames of the train split, same sensor). Earlier post-processing settings (fusion weights and IoU, the height-refine percentile, the pedestrian shifts and the choice of rule R6) were compared on 292 val frames of both infrastructure sensors.

  2. Constants fitted on val_i1 labels.

    • The thresholds of score rules X1, X3 and X8.
    • The truck size step (+0.67 m length, +0.12 m bottom): medians of label minus box.
    • The vehicle geometry step V-G: length βˆ’0.125 m and width +0.067 m are medians of label minus box; height +0.18 m and z +0.06 m come from val_i1 height-error checks (vehicle height error βˆ’0.21 m and βˆ’0.15 m in two range bands; median bottom error +0.03 m). The V-G idea came from comparing our per-class board results with those of a public entry.
    • The pedestrian height offset P-Z (+0.03 m), chosen among nine settings on val_i1 and then read once on fresh.
  3. Unlabelled real points. 128 frames of the V2X-Real test split (sensor OS1-64; not the challenge's test frames) were used as points only, with no labels read. They give:

    • the static-background map P3: voxels occupied in at least half of the frames; it lowers the score of a test pedestrian box Γ—0.3 when most of its points fall in the map;
    • the targets of the realism calibration of our 50 synthetic frames;
    • the training targets of the point refiner R2, which moves the points of those 50 frames.

    64 of these 128 frames come from four clips whose names share a timestamp and scenario prefix with clips that hold forbidden frames; we have not established whether they are the same recording.

  4. Read, not shipped. A negatives model (neg-G) resized its synthetic walker boxes to the pedestrian label sizes of 109 frames of the same test-split set (1,749 labels); it failed the gate and is not in the final package.

Validation, model selection and realism calibration are on the organisers' list of allowed real-data uses. P3, R2's training on real points and the constants fitted on val_i1 are not on that list as such; we list them here so that they can be judged.

Licence

  • Weights and files here: Creative Commons Attribution-NonCommercial 4.0 (CC BY-NC 4.0).
  • Code: Apache License 2.0, in the GitHub repository.
  • Data: the challenge datasets are not redistributed. Each keeps its own licence.

Citation

Please cite the challenge's papers and datasets:

@article{shahbaz2026high,
  title={High-fidelity digital twins for bridging the sim2real gap in {LiDAR}-based {ITS} perception},
  author={Shahbaz, Muhammad and Agarwal, Shaurya},
  journal={IEEE Transactions on Intelligent Transportation Systems},
  volume={27},
  number={7},
  pages={8698--8709},
  year={2026},
  publisher={IEEE}
}

@article{shahbaz2026urbantwin,
  title={UrbanTwin: Synthetic Roadside LiDAR Datasets},
  author={Shahbaz, Muhammad and Agarwal, Shaurya},
  journal={IEEE Open Journal of Intelligent Transportation Systems},
  volume={7},
  pages={353--364},
  year={2026},
  publisher={IEEE}
}

@article{shahbaz2025urbantwin,
  title={UrbanTwin: Building High-Fidelity Digital Twins for Sim2Real LiDAR Perception and Evaluation},
  author={Shahbaz, Muhammad and Agarwal, Shaurya},
  journal={arXiv preprint arXiv:2509.02903},
  year={2025}
}
Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Paper for Sadbodh/urbantwin-sim2real-lidar