GeoScaffold: From Latent to Realized Geometry in Collaborative Detection

GeoScaffold teaser: NMS often discards the better-localized collaborative box; GeoScaffold improves the box at the top-scored anchor

Code, Paper (arXiv, coming soon)

Kang Yang1   Yance Fang1   Gaochen Cui2   Xiaoqiang Jin3   Deying Li1   Yongcai Wang1
1Renmin University of China    2CFINS, Tsinghua University    3QiYuanLab

News

  • [2026-11] Checkpoints for the UrbanIng-V2X benchmark are released: the Native and +GeoScaffold models of 11 fusion methods, and the Ego-only, Early Fusion, and Late Fusion reference detectors.

GeoScaffold is a training-only auxiliary loss for collaborative 3D detectors. At each positive anchor, it predicts which cells of a grid over the anchor lie inside the ground-truth box, and this dense supervision trains the detector toward better-localized boxes. The head is removed at inference, so each detector keeps its architecture, communication, and inference cost. Across 11 fusion methods on the UrbanIng-V2X validation split, GeoScaffold raises AP@0.5 by 0.98 and AP@0.7 by 0.55 points on average.

This repository holds the checkpoints of the GeoScaffold code. Each folder under checkpoints/urbaning/ contains the final checkpoint of one run and the config of that run, and the code reads both from the folder passed to --model_dir.

License/Terms of Use

The checkpoints are released under the UCLA Academic Software License of the code base (LICENSE), which covers academic and non-profit research use. They were trained on the UrbanIng-V2X dataset, which its authors release under CC BY-NC-ND 4.0.

Checkpoints

UrbanIng-V2X validation split (crossing1/2), AP in percent, final checkpoints. Each Native/+GeoScaffold pair shares data, seed, sample order, augmentation, optimizer settings, learning-rate schedule, and training length.

Method Arm Folder File Size (MB) Evaluation flags AP@0.3 / 0.5 / 0.7
Ego-only reference checkpoints/urbaning/ego_only/native/ net_epoch20.pth 25.3 --fusion_method nofusion --gt_config configs/urbaning/fcooper/native.yaml 22.55 / 16.12 / 5.41
Early Fusion reference checkpoints/urbaning/early_fusion/native/ net_epoch20.pth 25.3 --fusion_method early 23.76 / 17.35 / 6.46
Late Fusion reference checkpoints/urbaning/late_fusion/native/ net_epoch20.pth 25.3 --fusion_method late 27.25 / 19.40 / 6.88
AttFuse Native checkpoints/urbaning/attfuse/native/ net_epoch20.pth 27.9 --fusion_method intermediate 24.25 / 18.18 / 7.16
AttFuse +GeoScaffold checkpoints/urbaning/attfuse/geoscaffold/ net_epoch20.pth 31.6 --fusion_method intermediate 25.19 / 18.89 / 7.30
F-Cooper Native checkpoints/urbaning/fcooper/native/ net_epoch60.pth 27.9 --fusion_method intermediate 31.62 / 25.35 / 11.19
F-Cooper +GeoScaffold checkpoints/urbaning/fcooper/geoscaffold/ net_epoch60.pth 31.6 --fusion_method intermediate 33.12 / 27.43 / 12.57
CoBEVT Native checkpoints/urbaning/cobevt/native/ net_epoch20.pth 37.4 --fusion_method intermediate 28.27 / 22.45 / 8.76
CoBEVT +GeoScaffold checkpoints/urbaning/cobevt/geoscaffold/ net_epoch20.pth 41.1 --fusion_method intermediate 28.53 / 22.94 / 9.07
Where2comm Native checkpoints/urbaning/where2comm/native/ net_epoch60.pth 27.9 --fusion_method intermediate 30.21 / 24.89 / 11.13
Where2comm +GeoScaffold checkpoints/urbaning/where2comm/geoscaffold/ net_epoch60.pth 31.6 --fusion_method intermediate 30.90 / 26.16 / 11.78
V2X-ViT Native checkpoints/urbaning/v2xvit/native/ net_epoch30.pth 38.3 --fusion_method intermediate 33.86 / 26.38 / 11.16
V2X-ViT +GeoScaffold checkpoints/urbaning/v2xvit/geoscaffold/ net_epoch30.pth 42.0 --fusion_method intermediate 36.35 / 29.08 / 12.58
V2VNet Native checkpoints/urbaning/v2vnet/native/ net_epoch20.pth 52.9 --fusion_method intermediate 22.84 / 17.11 / 6.65
V2VNet +GeoScaffold checkpoints/urbaning/v2vnet/geoscaffold/ net_epoch20.pth 56.7 --fusion_method intermediate 23.48 / 17.95 / 7.07
DiscoNet Native checkpoints/urbaning/disconet/native/ net_epoch40.pth 81.1 --fusion_method intermediate 30.64 / 25.52 / 11.37
DiscoNet +GeoScaffold checkpoints/urbaning/disconet/geoscaffold/ net_epoch40.pth 84.9 --fusion_method intermediate 30.84 / 25.75 / 11.61
CoAlign Native checkpoints/urbaning/coalign/native/ net_epoch30.pth 49.4 --fusion_method intermediate 28.22 / 19.96 / 7.38
CoAlign +GeoScaffold checkpoints/urbaning/coalign/geoscaffold/ net_epoch30.pth 53.2 --fusion_method intermediate 28.27 / 20.10 / 7.45
ERMVP Native checkpoints/urbaning/ermvp/native/ net_epoch30.pth 32.0 --fusion_method intermediate 29.19 / 23.76 / 10.04
ERMVP +GeoScaffold checkpoints/urbaning/ermvp/geoscaffold/ net_epoch30.pth 35.8 --fusion_method intermediate 29.86 / 24.15 / 10.72
HEAL Native checkpoints/urbaning/heal/native/ net_epoch30.pth 20.9 --fusion_method intermediate 26.59 / 20.72 / 9.13
HEAL +GeoScaffold checkpoints/urbaning/heal/geoscaffold/ net_epoch30.pth 24.6 --fusion_method intermediate 28.12 / 22.53 / 9.41
S-AdaFusion Native checkpoints/urbaning/sadafusion/native/ net_epoch30.pth 27.9 --fusion_method intermediate 29.56 / 21.43 / 7.61
S-AdaFusion +GeoScaffold checkpoints/urbaning/sadafusion/geoscaffold/ net_epoch30.pth 31.6 --fusion_method intermediate 29.99 / 21.49 / 7.99

25 checkpoints, 964 MB in total. The CoAlign alignment files (checkpoints/urbaning/coalign/alignment/) add 139 MB.

checkpoints/urbaning/coalign/alignment/ holds train_all_agents.json and val_all_agents.json, the stage-1 box-alignment results that CoAlign reads during training and evaluation.

Usage

Download the checkpoints into the root of the code repository, check that every model loads, and evaluate a checkpoint with the flags listed in the table:

git clone https://github.com/sidiangongyuan/GeoScaffold.git
cd GeoScaffold
hf download sidiangongyuan/GeoScaffold --local-dir . --include "checkpoints/*"
python tools/check_checkpoints.py
python -m opencood.tools.inference --model_dir checkpoints/urbaning/v2xvit/geoscaffold --fusion_method intermediate --gpu cuda:0

The code README covers installation, data preparation, evaluation, and training.

Citation

The BibTeX entry will be added with the arXiv version.

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. ๐Ÿ™‹ Ask for provider support