YAML Metadata Warning:The pipeline tag "3d-object-detection" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, audio-text-to-text, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-ranking, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, image-text-to-image, image-text-to-video, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, video-text-to-text, keypoint-detection, visual-document-retrieval, any-to-any, video-to-video, other
UrbanTwin LUMPI Track — Ensemble Detectors
3rd Place Winner | Combined Score: 0.4637 | ECCV 2026 DriveX Workshop
🏆 Overview
Four-model ensemble trained on synthetic CARLA LiDAR data for sim-to-real 3D object detection on the LUMPI (Leibniz University Multi-Perspective Intersection) dataset.
Model Ensemble
| Model | Architecture | Training Data | Epoch | Size | Role |
|---|---|---|---|---|---|
| v1 | PointPillar | lumpi_v1 (archive) | 30 | 56 MB | Diversity + robustness |
| v4 | PointPillar | lumpi (range-rescoped) | 30 | 56 MB | Baseline detection |
| v3 | PointPillar | lumpi_v3 (geometry-rescaled) | 30 | 56 MB | Primary model ⭐ |
| v5 | SECOND (voxel) | lumpi (range-rescoped) | 38 | 62 MB | Architecture diversity |
📊 Performance
| Metric | Value | Split |
|---|---|---|
| 3D mAP @ IoU 0.5 | 0.1183 | 50 held-out test frames |
| Realism Score | 0.9057 | CD/MMD/EMD/FPD (synthetic vs. real) |
| Combined Score | 0.4637 | 0.6 × detection + 0.4 × realism |
| Ranking | 3rd Place | LUMPI Track |
Per-Class Results
| Class | AP@0.5 | GT Count | Predictions |
|---|---|---|---|
| Car | 0.6112 | 1,258 | 5,964 |
| Person | 0.1546 | 512 | 7,641 |
| Truck | 0.0255 | 192 | 1,511 |
| Bicycle | 0.0231 | 82 | 4,229 |
| Van | 0.0130 | 79 | 2,110 |
| Bus | 0.0000 | 25 | 459 |
| Motorcycle | 0.0006 | 27 | 33,912 |
| Mean (7-class) | 0.1183 | 2,175 | 55,826 |
🔧 Model Architecture
PointPillar & SECOND from OpenPCDet:
- Pillar-based & voxel-based 3D detection
- Trained on 64-channel LiDAR point clouds
- Class-agnostic pre-NMS, per-class post-processing
📥 Usage
import torch
from pcdet.models import build_network
from pcdet.config import cfg_from_yaml_file
# Load configuration
cfg_from_yaml_file("cfgs/pointpillar_lumpi_v3.yaml", cfg)
# Build model
model = build_network(model_cfg=cfg.MODEL, num_class=8, dataset=None)
# Load checkpoint (v3 is strongest single model)
state = torch.load("v3_pointpillar_rescaled_epoch30.pth")
model.load_state_dict(state["model_state"])
model.cuda().eval()
# Inference on point clouds
with torch.no_grad():
predictions = model(data_dict)
For ensemble inference, use build_g.py from the GitHub repository.
🎯 Key Contributions
- Geometry Rescaling: Synthetic object dimensions rescaled to match real-world medians (+0.009 mAP)
- Per-Class Weighted Ensemble: Learned weights favor v3 for Person/Car (+0.018 mAP)
- Domain-Adaptive Calibration: Corrects size/height biases (+0.042 mAP — largest gain)
- p01 Point-Count Filter: Removes noise detections (+0.044 mAP)
📖 Citation
If you use these models or the code, please cite:
@techreport{urbantwin2026lumpi,
title={Ensemble Detectors with Domain-Adaptive Post-Processing for Sim2Real LiDAR Object Detection},
author={Anonymous Team},
year={2026},
note={3rd Place, UrbanTwin LUMPI Track, ECCV 2026 DriveX Workshop},
url={https://huggingface.co/Nithiishaa/urbantwin-lumpi-checkpoints}
}
🔗 Resources
- Full Challenge Report: CHALLENGE_REPORT.md
- Reproduction Guide: REPRODUCE_G.md
- GitHub Repository: urbantwin-lumpi
- Challenge: UrbanTwin Sim2Real LiDAR Challenge
⚖️ License
MIT License — See LICENSE
📝 Notes
- All models trained exclusively on synthetic CARLA data
- No real LiDAR labels used for detector weights
- Post-processing (calibration, filtering) fitted on public 217-frame real train split
- Realism metrics evaluated on 50 held-out real reference frames
Model Uploaded: September 2026
Framework: PyTorch 2.0.1 + OpenPCDet
CUDA: 11.8+ (inference CPU-compatible)