DriveX Workshop, ECCV 2026  ·  Oral Presentation

MAGneT-3D: Monocular and Domain-Generalizable Temporal 3D Detection

1TU Munich 2TU Darmstadt 3ETH Zurich 4University of Cambridge 5MCML 6ELIZA 7DeepScenario

†Equal contribution

Left: a Venn diagram of monocular, temporal and domain-generalization methods, with MAGneT-3D alone in the triple intersection. Right: a radar chart of nuScenes Detection Score on the source dataset and three unseen target datasets, where MAGneT-3D encloses every baseline.
MAGneT-3D overview. MAGneT-3D bridges monocular, temporal modeling and domain generalization (left) for strong cross-dataset 3D detection (right). Prior detectors instead treat these settings as distinct or paired, leaving the intersection unexplored. On our new cross-dataset benchmark, MAGneT-3D trained on nuScenes outperforms on both in-domain (source) and all unseen (target) datasets. We report the nuScenes Detection Score (NDS, in %, ↑).

Abstract

Monocular temporal 3D detection aims to detect objects in 3D, given a monocular video. Query-based 3D detectors unify detection and cross-view association, but their learnable queries fit to the spatial distribution of the training data (e.g., field-of-view). We show that this issue is especially severe when these models are applied to monocular video, hindering generalization to unseen datasets and environments. To address this limitation, we introduce MAGneT-3D, the first method for domain-generalized monocular temporal 3D object detection. Instead of relying on static learnable queries, we propose a Domain-Robust Anchor Generation (DRAG) approach that adaptively derives 3D proposals during inference. To further enable domain generalization, we propose a Temporal Refinement and Identity Merging (TRIM) strategy, reducing dependence on specific 3D proposals. To enable comprehensive domain-generalization evaluation, we establish a cross-dataset benchmark spanning nuScenes, Waymo, Lyft, and ONCE. Under zero-shot domain shifts, MAGneT-3D outperforms all baselines, improving NDS from 12.1 % to 18.6 % while also increasing in-domain accuracy.

Method

MAGneT-3D architecture: the current frame goes through a backbone into the Domain-Robust Anchor Generator, then spatial interaction and a temporal prediction refiner with tracking queries, feeding classification, box and re-identification heads. Insets show soft assignment during training and clustering-based deduplication at inference.
MAGneT-3D. Instead of static, learnable queries, we generate dynamic proposals (left top) based on image features (DRAG), eliminating source-domain biases. The proposals interact spatially with the image features and are fed, alongside historic queries, into the temporal prediction refiner. During training, our soft assigner matches each target to multiple proposals for dense supervision. During inference, duplicate objects are merged via clustering (TRIM).

Query-based detectors represent objects as a fixed set of learnable 3D queries. Training pulls those query reference points toward the locations where the source dataset happens to put its objects, and they are then frozen at inference. They cannot relocate to cover regions outside the source range, so predictions stay confined to the source angular distribution no matter how different the target camera is. MAGneT-3D builds on StreamPETR and replaces the static query set with two components.

Domain-Robust Anchor Generator (DRAG). An FCOS3D head attached to the backbone predicts a projected 2D centre and a depth per frame, which are back-projected with the current intrinsics to form 3D anchors. Two adjustments make this camera-agnostic: the network regresses a focal-normalized virtual depth rather than metric depth, and feature-pyramid level assignment is performed in focal-normalized space against a fixed reference focal length, so an instance always routes to the same level regardless of intrinsics.

Temporal Refinement and Identity Merging (TRIM). Because DRAG’s anchors already lie close to the targets, one-to-one Hungarian matching only ever teaches the refiner to make trivial adjustments — the wrong skill for an unseen domain, where anchors are worse. We instead match each target to several nearby proposals for dense supervision, and train a re-identification embedding with a supervised contrastive loss. At inference, predictions are clustered by that embedding and only the highest-confidence one per cluster is kept, recovering the precision dense supervision costs.

Results

Trained on Metric
Method mAP ↑ NDS ↑
nuScenes Waymo Lyft ONCE Avg. Cross nuScenes Waymo Lyft ONCE Avg. Cross
StreamPETR 26.5 4.8 5.0 8.2 6.0 28.5 13.1 10.6 12.7 12.1
Sparse4D v3 24.8 4.4 8.6 11.1 8.0 27.0 9.5 13.1 14.7 12.4
Far3D 28.3 2.3 5.6 9.3 5.7 29.8 10.8 12.8 13.2 12.3
BEVFormer v2 21.8 6.2 2.5 10.6 6.4 22.2 11.2 9.1 14.5 11.6
MAGneT-3D (ours) 28.5 10.1 14.3 12.4 12.3 32.9 17.2 19.5 19.1 18.6

Comparison against state of the art. Each model is trained on a single source dataset and evaluated on all target datasets; Avg. Cross is the mean over the three unseen datasets. In-domain results are greyed and excluded from Avg. Cross. mAP and NDS in % (↑); best per column within the shown training source in bold. Relative to the best baseline, MAGneT-3D improves average cross-dataset NDS by 50 % on nuScenes, 83 % on Waymo and 16 % on Lyft.

Four driving scenes with corresponding bird's-eye-view plots. MAGneT-3D boxes in blue align closely with green ground truth boxes, while StreamPETR in orange produces many spurious detections. The rightmost column shows a nuScenes failure case.
Qualitative results. MAGneT-3D yields fewer false positives and more accurate predictions than StreamPETR. All models are trained on nuScenes and evaluated across nuScenes, Lyft, and Waymo. Best viewed in color; zoom in for details. BEV color coding: Ground truth , StreamPETR , and MAGneT-3D .

Ablations

Ablate Metric
Configuration DRAG TRIM mAP ↑ NDS ↑
nuScenes Waymo Lyft ONCE Avg. Cross nuScenes Waymo Lyft ONCE Avg. Cross
1. StreamPETR (baseline) – – 26.5 4.8 5.0 8.2 6.0 28.5 13.1 10.6 12.7 12.1
2. + DRAG ✓ – 20.0 7.9 8.6 8.8 8.4 25.6 16.3 13.9 13.9 14.7
3. + TRIM – ✓ 27.1 6.9 8.6 11.9 9.1 27.0 11.7 13.6 16.4 13.9
4. MAGneT-3D ✓ ✓ 28.5 10.1 14.3 12.4 12.3 32.9 17.2 19.5 19.1 18.6

Main ablation. Neither component suffices alone: DRAG alone raises average cross-dataset NDS to 14.7 but lowers in-domain accuracy to 25.6, and TRIM alone reaches 13.9. Together they reach 18.6 cross-dataset while also raising in-domain accuracy to 32.9.

BibTeX

@inproceedings{kotb2026magnet3d,
  title     = {MAGneT-3D: Monocular and Domain-Generalizable Temporal 3D Detection},
  author    = {Kotb, Mohamed and Meier, Johannes and Reich, Christoph and
               Dhaouadi, Oussema and Denninger, Luis and Cremers, Daniel},
  booktitle = {European Conference on Computer Vision (ECCV) Workshops},
  year      = {2026}
}