acceptodds
Withdrawn from ICLR 2027

AnchorFuse: Robust Anchor-Centered Temporal Fusion for Multimodal Aerial-Ground Place Recognition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal aerial-ground place recognition localizes a ground platform by matching camera–LiDAR observations to georeferenced aerial imagery. However, descriptor-level temporal aggregation does not explicitly distinguish evidence relevant to the queried location from observations of neighboring locations, and retrieval can degrade when context is unavailable. We propose AnchorFuse, an anchor-centered framework that explicitly separates the retrieval target from its temporal context. An anchor observation defines both the positive aerial references and the coordinate frame for fusion, keeping the retrieval target fixed as auxiliary observations change. Building on this formulation, Anchor-Aligned Cross-Frame Fusion (AACF) transforms local camera–LiDAR features into the anchor frame and selects nearby evidence according to spatial proximity, temporal distance, and learned reliability. A gated residual integrates the selected evidence into the query descriptor. To accommodate missing context, Robust Anchor-Memory Consistency (RAMC) jointly supervises clean and corrupted queries against the same retrieval target and aligns corrupted descriptors with a stop-gradient clean target. Context corruption preserves the multimodal anchor, and RAMC adds no inference-time branch. AnchorFuse achieves R@1 of 44.28% on KITTI360-AG and 78.35% on nuScenes-AG. On KITTI360-AG, RAMC improves R@1 over AACF by 17.03 percentage points when two non-anchor frames are removed, compared with 1.49 points on clean queries.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Aerial-Ground Place Recognition, Cross-View Geo-Localization, Multimodal Retrieval, Temporal Fusion, Robust Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_yvazOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript introduces AnchorFuse for multimodal aerial-ground place recognition using temporal camera–LiDAR observations. AACF aligns local features to an anchor frame and refines the query descriptor through sparse readout and a gated residual. RAMC combines retrieval supervision on queries with missing context and consistency to a stop-gradient clean descriptor, while preserving the anchor observation.

Critical strengths

The problem of using temporal context while keeping the retrieval target fixed is relevant, and incorporating local geometry into temporal fusion is reasonable. The experiments report improvements on two benchmarks, and several tables (Tabs. 7, 10, and 11) well provided useful ablations of the AACF components, RAMC objectives, and auxiliary losses. This manuscript also reports repeated runs, context-corruption experiments, and inference cost, while acknowledging the smaller robustness gains on nuScenes-AG.

Critical weaknesses

I'm not an expert on this research area, so i just write few concerns on my perspective, as follows:

  • Their proposed CVLNet already uses query-centered geometric fusion, while Table 7 lacks a matched control that restricts local memory to anchor entries. I think that comparisons with these alternatives would help isolate the benefit of the proposed cross-frame readout.

  • Because the base descriptor uses the full temporal window well, the gate in Eq. (8) does not provide an anchor-only fallback. Even RAMC improves over AACF, removing all context cameras yields only 14.59% R@1 on KITTI360-AG versus 35.06% for the single-observation baseline, making a fallback comparison important.

  • Table 1 uses three observations for GSPR and SeqMamba-MPR but five for AnchorFuse, while Appendix A.2 omits sequential dropout from CVIS/SeqGeo. Comparisons should match temporal inputs, retain the baselines’ key training mechanisms, and include the same context-corruption conditions.

Questions for authors

Please check the former section, weakness section and carefully address all the concerns.

Reviewer_FWLgOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes AnchorFuse, an anchor-centered framework for multimodal aerial-ground place recognition that leverages temporal camera–LiDAR observations to improve cross-view retrieval. The framework introduces AACF to select and integrate spatially relevant features from neighboring observations and RAMC to enhance retrieval robustness under missing temporal and sensor context.

Critical strengths
  • The paper addresses a practical problem in multimodal aerial-ground place recognition by incorporating temporal camera–LiDAR observations.
  • The proposed anchor-centered framework combines geometric alignment, local feature selection, and gated descriptor refinement, providing a structured approach to integrating temporal information.
Critical weaknesses
  1. The methodological novelty of the proposed framework could be further highlighted. AACF combines several established techniques, including pose-based alignment, local feature selection, and gated residual fusion, while RAMC builds on consistency learning under context corruption. The authors should better distinguish the key methodological contributions from a modular combination of existing techniques.
  2. The paper argues that descriptor-level temporal aggregation introduces ambiguity between the queried location and neighboring observations. However, conventional temporal retrieval methods can also explicitly designate an observation as the retrieval target.
  3. Comparisons with stronger temporal attention-based or pose-aware fusion methods using the same five-frame RGB–LiDAR input would provide a more convincing evaluation of the proposed AACF design.
  4. The effectiveness of RAMC varies considerably across benchmarks. According to Table A9, RAMC improves R@1 by only 0.87 percentage points on nuScenes-AG when two context frames are removed, compared with 17.03 points on KITTI360-AG. Table 2 also shows a limited improvement on clean nuScenes-AG queries. These results raise questions about the effectiveness and generalizability of RAMC across different datasets.
  5. The proposed method relies on future observations within a symmetric temporal window. As shown in Table 4, using past-only observations leads to a noticeable performance degradation. Does this dependence on future observations limit the applicability of AnchorFuse to online localization scenarios?
  6. Several methodological descriptions could be improved. For example, Section 3.2 states that the base descriptor zaz_a already incorporates the entire temporal window, while AACF subsequently retrieves additional local evidence from the same observations to refine this descriptor. The distinction between the information captured by the base descriptor and the additional contribution of AACF should be more clearly explained.
Questions for authors
  1. Could conventional temporal fusion methods also learn to identify and emphasize information relevant to a designated retrieval target through end-to-end training? What additional benefits does the proposed anchor-centered formulation provide beyond such implicit learning?
  2. How can the authors further demonstrate the effectiveness of AACF compared with stronger temporal fusion methods under the same input settings?
  3. How well does AnchorFuse generalize across different datasets and practical deployment scenarios, considering the varying effectiveness of RAMC and the reliance on future observations for temporal fusion?
Reviewer_q1zKOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies multimodal aerial-ground place recognition using temporal camera-LiDAR observations. The authors define an explicit anchor to fix the retrieval target and fusion coordinates, and introduce Anchor-Aligned Cross-Frame Fusion (AACF) to align local features, select nearby evidence, and refine the query descriptor through gated fusion. Robust Anchor-Memory Consistency (RAMC) supervises both complete and frame-dropped queries and aligns corrupted descriptors with a stop-gradient clean target. Experiments on KITTI360-AG and nuScenes-AG compare temporal fusion methods and evaluate missing context, pose perturbations, and temporal outliers. AnchorFuse achieves R@1 of 44.28% and 78.35% on the two datasets. RAMC substantially improves robustness over AACF on KITTI360-AG, although several corrupted-query results remain below the single-observation baseline and improvements on nuScenes-AG are limited.

Critical strengths
  1. The explicit anchor clearly distinguishes the queried location from neighboring observations. Pose alignment and local evidence selection provide a well-motivated approach to temporal fusion.
  2. Matched-backbone comparisons show useful gains. On KITTI360-AG, AACF exceeds learned frame weighting by 3.86 R@1 points, while RAMC improves two-frame-removal performance from 15.38% to 32.41% without an additional inference branch.
  3. The experiments include objective ablations, missing-context tests, pose and temporal-outlier perturbations, five-seed results, and zero-shot road-map evaluation. The paper also acknowledges future-frame delay and the limited robustness gains on nuScenes-AG.
Critical weaknesses
  1. The construction of the full-window base descriptor z_a is insufficiently specified. The absence of matched controls without AACF refinement and with anchor-only local memory also limits attribution of the cross-frame fusion gains.
  2. Robustness gains are mainly measured against AACF, while several missing-context results fall below the intact-anchor single-observation baseline. On KITTI360-AG, removing two context frames reduces R@1 to 32.41%, compared with 35.06% for the single-observation model. Simpler temporal controls under identical corruptions are also missing.
Questions for authors
  1. Could the authors specify the temporal aggregation and report matched controls without AACF refinement and with anchor-only local memory?
  2. How do zero-context AnchorFuse, the single-observation model, temporal averaging, and learned frame weighting perform under identical removals?
  3. What gains remain when temporal pooling receives the same frame dropout and clean-target consistency training?
  4. Does past-only AnchorFuse outperform matched past-only temporal baselines? Its KITTI360-AG R@1 is 35.72%, close to the 35.06% single-observation baseline, making the contribution of future context important to quantify.
  5. Could the authors report per-seed variability for the principal AACF and missing-context comparisons? Mean improvements alone do not establish the reliability of small gains, particularly on nuScenes-AG.

If your rebuttal clears things up for me, i'll consider raising the score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.