ArgoTweak: Towards Self-Updating HD Maps through Structured Priors

Conference: ICCV 2025 arXiv: 2509.08764 Code: https://KTH-RPL.github.io/ArgoTweak/ Area: Interpretability Keywords: HD Map, Map Updating, Change Detection, Dataset, Autonomous Driving

TL;DR

This paper proposes ArgoTweak, the first HD map dataset providing complete triplets of "prior map + current sensor data + up-to-date ground-truth map." It decomposes large-scale map modifications into element-level atomic changes via a bijective change mapping framework, and introduces interpretable evaluation metrics (mAPC/mACC). Models trained on ArgoTweak reduce the sim2real gap by more than 10× compared to synthetic-prior baselines.

Background & Motivation

Background: HD maps are a core component of autonomous driving, providing precise lane-level information. Recent work has shifted from manual annotation toward end-to-end online generation based on BEV features (e.g., MapTR, LaneSegNet). Some methods attempt to leverage prior maps to improve generation quality.

Limitations of Prior Work: - Lack of complete data: No public dataset simultaneously provides triplets of "prior map + current sensor data + up-to-date ground-truth map." Existing methods can only train with synthetic priors (scripted modifications, noise injection, random element deletion, etc.). - Large sim2real gap: Synthetic priors fail to capture the structured, semantically correlated nature of real-world changes (e.g., adding a bike lane entails modifying lane markings and connectivity), leading to significant performance degradation in real scenarios. - Insufficient evaluation metrics: Standard mAP cannot distinguish between "preserving unchanged regions" and "correctly updating changed regions." Models trained on different prior qualities may achieve nearly identical mAP scores (experimentally verified: ArgoTweak vs. synthetic priors differ by only ~1% mAP, despite large qualitative gaps).

Key Challenge: Achieving truly self-updating HD maps requires addressing deficiencies in data, models, and evaluation simultaneously.

Core Idea: Construct a real-prior dataset annotated with a bijective mapping framework, enabling models to learn interpretable element-level change detection and updating.

Method

Overall Architecture

ArgoTweak contributes across three dimensions — dataset, model, and metrics: - Dataset: Built upon the Argoverse 2 Map Change Dataset with manually annotated real prior maps, forming complete triplets. - Model: LaneSegNet backbone + map prior encoder + interpretable change assessment heads. - Metrics: mAPC (change-aware precision) + mACC (change detection accuracy).

Key Designs

1. Bijective Change Mapping:

   • Function: Systematically decomposes large-scale map modifications into traceable element-level atomic changes.
   • Mechanism: Defines an atomic change set \(\mathbf{A} = \{\)geometry, markings, type, connectivity, insertion, deletion\(\}\) and a structural update set \(\hat{\mathbf{Y}} = \{\)shape, appearance, function, lane graph, lane number\(\}\). A mapping between the two is constructed via surjectivity (every structural change can be explained by a combination of atomic changes) and injectivity (a given structural change cannot have two fundamentally different representations).
   • Ambiguity resolution: Disambiguation rules based on lane graph topology are introduced — when a change also modifies topology (adding/removing connections), it is represented as insertion/deletion; otherwise, in-place edits (geometry/marking/type) are used.
   • Design Motivation: Ensures annotation consistency and interpretability, enabling models to learn structured change patterns rather than random perturbations.

2. Dataset Construction:

   • Function: Constructs a training set (synthetic yet realistic priors) and a test set (real-world priors).
   • Mechanism: Training priors are generated by applying structured modifications within the bijective framework to Argoverse 2 ground-truth maps. The test set uses real outdated maps from the Argoverse 2 Map Change Dataset validation split, with re-annotated up-to-date ground truth.
   • Scale: 697 training / 102 validation / 111 test scenarios, average duration 56 s.
   • Additional processing: OpenLane-V2-style lane merging and unified crosswalk edge orientation.

3. Interpretable Prior-Assisted Map Network:

   • Function: Achieves interpretable map updating while maintaining modeling flexibility.
   • Mechanism: Uses LaneSegNet as the backbone and adds a prior encoder (10-point 2D coordinates × left/right/centerline + one-hot marking type). Priors are injected into BEV features via cross-attention. Change assessment operates at two levels: a primary head (multi-class: No Change / Insertion / Deletion / Other) and secondary heads (binary: geometry and marking changes), with mutually exclusive and co-occurring categories handled separately.
   • Loss: \(\mathcal{L} = \lambda_{\text{vec}}\mathcal{L}_{\text{vec}} + \lambda_{\text{seg}}\mathcal{L}_{\text{seg}} + \lambda_{\text{cls}}\mathcal{L}_{\text{cls}} + \lambda_{\text{type}}\mathcal{L}_{\text{type}} + \lambda_{\text{cd,prim}}\mathcal{L}_{\text{cd,prim}} + \sum_i \lambda_{\text{cd,sec}}^i \mathcal{L}_{\text{cd,sec}}^i\)

4. Change-Aware Dual Metrics (mAPC + mACC):

   • Function: Separately evaluate map stability (ability to preserve unchanged regions) and responsiveness (ability to update changed regions).
   • Mechanism:
     • mAPC: Requires that the predicted change state matches the ground-truth change state \(\hat{c}_V = c_V\) during prediction-to-ground-truth matching, computing AP per change category and averaging.
     • mACC: Computes binary detection accuracy per frame and per change type, averaging over accuracies for positive (changed) and negative (unchanged) cases.
   • Design Motivation: High mACC but low mAPC indicates that changes are detected but not accurately localized; low mACC indicates an overly conservative model. Plain mAP cannot capture these distinctions.

Loss & Training

• 10 epochs, batch size 8, AdamW optimizer, 8× NVIDIA A10G.
• ResNet-50 pretrained feature extractor, camera-only input.
• Map crop 50×50 m², ~4 FPS (single A10G).

Key Experimental Results

Main Results

mAP comparison under different priors (without change annotation):

Prior Type	AP_ls	AP_pc	mAP
No prior (baseline)	32.9	45.9	39.4
Continuous perturbation (noise)	71.6	75.5	73.5
Discrete modification (deletion/shift)	71.0	71.9	71.5
Rule-based editing	74.2	79.3	76.7
ArgoTweak	75.8	79.6	77.7

• The mAP gap is only ~1%, yet qualitative differences are substantial: the ArgoTweak model captures complex road updates, whereas synthetic-prior models can only make minor corrections or overfit to marking changes.

Sim2Real gap evaluation:

Training Prior	Δ mACC (Val→Test)
Rule-based editing	-36.0
ArgoTweak	-3.5

The sim2real gap is reduced by more than 10× when training on ArgoTweak.

Ablation Study

Comparison of annotation granularity:

Change Annotation	mAP	mAPC	mACC
None	77.7	-	-
Binary annotation (c/¬c)	77.5	66.5	70.5
Full atomic annotation	78.8	64.6	8.7*

*Note: mACC is more stringent under fine-grained annotation due to the need to distinguish multiple change types.

• Change annotations not only support fine-grained evaluation but also serve as auxiliary training signals (annotated models achieve higher mAP).
• Removing geometry change annotation improves mAPC (current map generation precision is insufficient to reliably distinguish subtle shape changes).

Personal Thoughts

• Highlights: The first dataset to provide complete "prior–sensor–ground-truth" triplets; the bijective mapping framework is formally elegant and practically useful. Experiments clearly demonstrate the inadequacy of mAP and the severity of the sim2real gap.
• Limitations: Training set priors are still manually modified (not real outdated maps); the test set is relatively small (111 scenarios); only camera input is used, without exploiting LiDAR.
• Insights: Interpretable change annotations serve not only as an evaluation tool but also as effective training signals. The "prior-assisted" map generation paradigm holds broad application potential.

Highlights & Insights

Limitations & Future Work

Related Work & Insights

Rating

• Novelty: TBD
• Experimental Thoroughness: TBD
• Writing Quality: TBD
• Value: TBD

Related Papers

• [ICCV 2025] CE-FAM: Concept-Based Explanation via Fusion of Activation Maps
• [ICCV 2025] AIM: Amending Inherent Interpretability via Self-Supervised Masking
• [ICCV 2025] VITAL: More Understandable Feature Visualization through Distribution Alignment and Relevant Information Flow
• [NeurIPS 2025] Latent Principle Discovery for Language Model Self-Improvement
• [NeurIPS 2025] Dataset Distillation for Pre-Trained Self-Supervised Vision Models