Pith. sign in

REVIEW 3 major objections 5 minor 17 references

PairUAV relative localization should give heading and range separate decoder readouts, not a single homogeneous head.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 18:30 UTC pith:J6GL6E2W

load-bearing objection Axis-conditioned readouts are a genuinely useful idea for PairUAV, but the headline claim rests on single-run local controls that need more seeds and an official full-depth baseline before I'd take it as robust. the 3 major comments →

arxiv 2607.17268 v1 pith:J6GL6E2W submitted 2026-07-19 cs.CV

PACE: Polar Axis-Conditioned Estimation for PairUAV Relative Localization

Ze Rong This is my paper
classification cs.CV
keywords UAV relative localizationpolar navigation commandheading and range estimationrelative pose regressionaxis-conditioned readoutdecoder evidence allocationcheckpoint disagreementrange error tail
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that PairUAV relative localization—mapping two UAV images to a heading angle and a range—should not treat the two outputs as homogeneous coordinates that consume the same decoder evidence and share the same optimization state. Controlled readout probes show heading improves when reading middle and late relational layers, while range stays reliable on a direct late metric path; the best checkpoints for the two axes disagree on 80.8% of a validation trajectory, and range errors have a distinct high-error tail. The proposed method, PACE, keeps one shared image-pair representation and gives each polar axis a separate readout interface, with MDHR and PAAER as two instantiations. If the paper is right, the design lesson is that multi-output pose decoders should allocate evidence per output coordinate rather than read one terminal feature.

Core claim

The central claim is that the map (I_a, I_b) to (θ, d) hides a design choice: sharing pair context does not force both axes to use identical decoder evidence. Matched readout-depth controls show mid–late features help heading while range benefits from a direct late metric interface; heading-best and range-best checkpoints disagree for 80.8% of pairs on one surface and 77.3% on an independent surface; and range errors exhibit a separate high-error tail. PACE instantiates the axis-conditioned view by keeping a shared encoder–decoder and attaching per-axis readouts: MDHR aggregates middle and late features only for heading while retaining a terminal range path, and PAAER uses an asymmetric prot

What carries the argument

The central object is the axis-conditioned readout interface over a shared encoder–decoder feature bank Z = {z(1),...,z(L)}. Instead of a homogeneous readout H(z(L)) predicting both axes, PACE selects a decoder subset S_a per axis: heading may aggregate middle and late layers (MDHR) or use query-bridge attention over them (PAAER), while range is kept attached to the late layer z(L). This mechanism carries the argument because it lets the paper attribute observed gains to evidence allocation per axis, not to separating the underlying pose factors.

Load-bearing premise

The load-bearing premise is that the matched readout-depth controls and checkpoint-disagreement statistics, measured on fixed local surfaces, remain representative of the official hidden-test distribution; if the mid/late-versus-late ordering reverses on the full training budget or on the official distribution, the PACE design principle loses its empirical footing.

What would settle it

A controlled experiment would settle the claim: train the same shared backbone with matched budget and initialization but force readouts to late-only, mid–late, and full-depth, then evaluate on a fresh held-out distribution stratified by range magnitude. If late-only range or mid–late heading no longer dominates its alternatives—or if per-axis best-checkpoint disagreement drops to chance on a new trajectory—the axis-conditioned evidence allocation is refuted.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, joint relative-pose decoders should allocate decoder evidence per output coordinate rather than read one terminal feature for all outputs.
  • Heading and range can share the same image-pair representation without latent-factor disentanglement, so multi-task sharing remains worthwhile even when readout paths differ.
  • Range-tail errors are partly repairable: a tail-weighted continuation cut the p95, p99, and maximum range errors by 85.3%, 97.1%, and 95.2% in the paper's diagnostics.
  • The deterministic packaging (range fusion, 2-degree heading lattice, range support projection) is separate from learned estimation and should be revalidated whenever the output space changes.
  • The strongest raw predictor scores 0.002460 and the final package 0.001874 on the official hidden test, with PAAER contributing a slightly lower angle error than MDHR.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper leaves implicit is per-axis optimization control: since heading-best and range-best checkpoints disagree so often, per-axis early stopping or separate schedules could improve either coordinate without architectural change.
  • The range high-error tail may be amplified by the evaluator's relative-error denominators, so a range-stratified error report would test whether the tail is visual evidence or a denominator artifact.
  • The axis-conditioned readout principle generalizes beyond UAV localization: any multi-output regression sharing one representation (for example depth-plus-surface-normal prediction) could profit from checking whether different outputs read different decoder depths.
  • The fusion coefficients were tuned on aggregate leaderboard feedback, so the cleanest next test of the learned claim is to replace that fixed fusion with a trainable per-axis reliability head.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper addresses PairUAV relative localization, where a model maps two UAV images to a polar navigation command (heading and range). The central claim is that although heading and range share the same pairwise image context, they should not necessarily consume the same decoder evidence or optimization state. The authors propose PACE, which keeps a shared Reloc3r-style encoder–decoder but assigns axis-specific readout depths: heading may read middle and late features, while range remains attached to the late metric path. Two instantiations are presented: MDHR and PAAER. Evidence includes matched readout-depth controls on local surfaces (Table 1), checkpoint-disagreement statistics (Section 4.3), range-tail diagnostics (Table 2), and official hidden-test results (Table 3). The final submitted score of 0.001874 is explicitly separated from raw learned predictors and is obtained by fusing three range predictors with coefficients selected from aggregate leaderboard feedback plus deterministic lattice rounding and support projection.

Significance. If the central claim is robust, the paper makes a useful and potentially generalizable point: for coupled output coordinates that share a pairwise representation, the decoder evidence and optimization objective need not be homogeneous. The paper is unusually honest in separating raw learned predictors from deterministic challenge packaging, and it releases code, checkpoints, predictions, and reconstruction tools. The claims are falsifiable and the evaluation includes an external hidden test, which is a strength. However, the load-bearing architectural claim currently rests on a single matched control block with one run per condition, and the official hidden-test numbers do not include a matched homogeneous readout baseline. The paper's scientific contribution would be substantially strengthened by multi-seed comparisons and an official-test full-depth control.

major comments (3)
  1. [Section 4.2, Table 1] The central architectural claim—that mid/late readout for heading and late-only for range outperforms homogeneous full-depth readout—rests entirely on Table 1, and every row is a single run with no seed variance or confidence interval. At the one-epoch continuation, MDHR's proxy is 0.009866 versus 0.010497 for full-depth, a margin of about 6.4%; at 2.5k steps, the margins between several configurations are similarly small (MDHR 0.013425 vs. mid-only 0.013609). In deep fine-tuning experiments, run-to-run variance of this magnitude is common and could easily reverse the ordering. Please provide multiple seeds (at least 3–5) with mean/std or paired significance tests for the matched depth controls, and report the same for the capacity control and PAAER interface control.
  2. [Section 4.4, Table 3] The official hidden-test table compares MDHR and PAAER against Reloc3r and metric-aware RPR baselines, but it does not include a matched full-depth (early+middle+late) or late-only homogeneous readout baseline on the official test distribution. Thus the official numbers support the overall PACE pipeline but not the specific readout-depth allocation claim. The depth-specific claim remains supported only by the single-run local controls in Table 1. Please add official hidden-test raw predictors for the homogeneous full-depth and late-only readouts, trained under the same budget, initialization, and data protocol, so that the central architectural claim is tested on the same distribution as the headline results.
  3. [Section 4.4, Table 3 and Section 3.4] The paper states that PAAER has a 'slightly lower angle relative error' than MDHR (0.002636 vs. 0.002646), but this difference is 0.000010, about 0.4%, and is a single-run point estimate on the hidden test with no error bars. Given that the final package selects PAAER as the heading source and uses leaderboard-tuned fusion coefficients, this near-tie is being used to justify a design choice without statistical support. Please either provide uncertainty estimates (e.g., variance across seeds or bootstrap over public validation) or soften the claim to note that the angle difference is not distinguishable from noise.
minor comments (5)
  1. [Throughout] There are numerous formatting/typo issues: 'PairUA V', 'UA VM', missing spaces around arrows (e.g., '(I a, Ib)7→'), and 'the final score averages 1 2'. Please run a copyediting pass.
  2. [Section 3.1, Eq. (1)] The angle relative error uses |θ_i| in the denominator; for near-zero target headings this can become unstable. Please clarify the official rule for small-but-nonzero denominators, and consider reporting how many test pairs are affected.
  3. [Table 2] The table caption and text refer to 'PAAER 50k', 'MDHR 50k', and 'PAAER + tail 10k'. Please clarify in the caption that the third row is a continuation of the PAAER line, not a same-budget architecture, to avoid reader confusion.
  4. [Section 4.2, 'Capacity control'] The capacity-control result (heading/range MAE 1.7925/4.5583 to 1.8504/4.6857) is reported in prose without a table or error bars. Please move it into a table or otherwise make it as transparent as the main depth controls.
  5. [Section 5] The limitations section is candid and fair. One suggestion: explicitly state that the 80.8%/77.3% checkpoint-disagreement figures are label-assisted diagnostics and do not imply a deployable selection rule; Section 4.3 does say this, but it could be restated in the limitations for emphasis.

Circularity Check

0 steps flagged

No significant circularity: central architectural claims are checked against external hidden-test results and matched local controls; only disclosed leaderboard-tuned packaging is self-referential, and it is not load-bearing for the learned estimation claim.

full rationale

The paper's central claim — that heading and range benefit from different decoder readout depths within a shared Reloc3r-style representation — is supported by matched readout-depth controls in Table 1, which compare mid-late, late-only, early-late, and full-depth readouts under the same data surface, initialization, optimizer, and loss. The chosen MDHR readout is then evaluated on the external official hidden test (Table 3), alongside Reloc3r baselines, so the learned-estimator comparison is not defined by the paper's own outputs. The Reloc3r host is cited as external prior work, and there is no load-bearing self-citation, imported uniqueness theorem, or ansatz smuggled in via citation. The only self-referential element is Eq. (5): the fusion coefficients were selected using aggregate leaderboard feedback, and the resulting final package score of 0.001874 is explicitly labeled as deterministic challenge packaging, not a learned module or a general fusion law. This is disclosed in Section 3.4 and again in the Limitations. It is a test-set-tuning / robustness caveat rather than a circular derivation, because the raw MDHR and PAAER results and the matched architectural controls do not depend on those coefficients. Score 1 reflects this minor disclosed tuning element; no circularity pattern reduces the central derivation to its inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The paper's central claim is empirical; the ledger entries are the hand-chosen hyperparameters, the leaderboard-tuned fusion coefficients, and the distribution assumptions connecting local surfaces to the hidden test. No new physical entities or additional mathematical axioms are introduced.

free parameters (4)
  • Range fusion coefficients = w_P=0.511, w_M=0.189, w_B=0.300
    Selected during challenge-time submission sweeps from aggregate leaderboard feedback; enter Eq. (5) and directly set the final packaged score.
  • Tail-aware sample weight parameters = threshold 80, width 40, max multiplier 3
    wi = 1 + 2 clip((|d|-80)/40, 0, 1); hand-chosen to emphasize long-range pairs in the PAAER continuation; the paper notes no validation labels define them.
  • Range-loss weight gamma_d = 0.05 (MDHR), 0.10 (PAAER)
    Hyperparameters in Eq. (4) that balance the range error terms; varied between runs.
  • Heading lattice rounding bin = 2 degrees
    Deterministic packaging in Eq. (6); a fixed discretization chosen to exploit the challenge output structure.
axioms (4)
  • domain assumption Reloc3r-512 checkpoints provide a suitable shared pair representation for PairUAV after full-model fine-tuning.
    Section 3.3 and Section 4.2 adopt Reloc3r as the host; all results depend on this host choice.
  • domain assumption Local validation surfaces (val811 and the 4,096-pair surface) are representative of the official hidden-test distribution.
    The 80.8% checkpoint-disagreement statistic and the readout-depth preferences are measured on these surfaces; if they do not carry over to the official hidden test, the architectural claim weakens.
  • domain assumption The official evaluator's relative-error definitions and zero-denominator exclusion rule are as stated in Eq. (1).
    All reported scores inherit this protocol; if the official rule differs, the comparisons change.
  • standard math Standard deep-learning convergence assumptions for AdamW/mixed-precision fine-tuning at batch size 4.
    Section 3.3; the paper relies on standard optimizer behavior and does not prove convergence.

pith-pipeline@v1.3.0-alltime-deepseek · 6127 in / 10989 out tokens · 100319 ms · 2026-08-01T18:30:13.351977+00:00 · methodology

0 comments
read the original abstract

PairUAV relative localization maps two UAV images to a polar navigation command. Although heading and range share the same pairwise pose context, treating them as homogeneous coordinates forces both outputs to use the same decoder evidence and optimization state. Controlled readout probes reveal a different structure: the two axes favor different decoder-depth combinations, their best checkpoints disagree on 80.8% of a validation trajectory, and range errors exhibit a distinct high-error tail. We introduce method, Polar Axis-Conditioned Estimation, which retains a shared Reloc3r-style pair representation while assigning axis-specific readout interfaces. Heading uses mid/late relational evidence, whereas range remains attached to a direct late metric path. On the official hidden test, the strongest released raw predictor scores 0.002460; the complementary PAAER predictor scores 0.002514 with a slightly lower angle error. Deterministic challenge packaging, reported separately from learned estimation, yields the final score of 0.001874. Code, checkpoints, predictions, and reconstruction tools are available at https://github.com/zerong7777-boop/PairUAV-PACE.

Figures

Figures reproduced from arXiv: 2607.17268 by Ze Rong.

Figure 1
Figure 1. Figure 1: Patch-occlusion sensitivity on four selected validation pairs. Heading and range can respond to different [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Per-axis best-checkpoint shares on val811. Heading and range select different checkpoints for 80.8% of pairs; [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

17 extracted references · 12 linked inside Pith

  1. [1]

    The 4th workshop on UA Vs in multimedia: Capturing the world from a new perspective

    Fabian Deuser, Yaxuan Li, Tingyu Wang, Yujiao Shi, Anna Bößendörfer, Shaofei Huang, Xiao Pan, Zhedong Zheng, and Roger Zimmermann. The 4th workshop on UA Vs in multimedia: Capturing the world from a new perspective. InProceedings of the 34th ACM International Conference on Multimedia Workshop, 2026

    2026

  2. [2]

    Last-meter precision navigation for UA Vs: A diffusion-refined aerial visual servoing approach, 2026

    Yaxuan Li, Jiarui Zeng, Shaofei Huang, and Zhedong Zheng. Last-meter precision navigation for UA Vs: A diffusion-refined aerial visual servoing approach, 2026. URLhttps://arxiv.org/abs/2607.04352. 6 PACEPREPRINT

    Pith/arXiv arXiv 2026

  3. [3]

    ACMMM26 Last-Meter Precision Navigation for UA Vs, 2026

    UA VM 2026 Workshop Organizers. ACMMM26 Last-Meter Precision Navigation for UA Vs, 2026. URLhttps: //github.com/spyflying/ACMMM2026Workshop-UAV. PairUA V task description and evaluation protocol

    2026

  4. [4]

    TransGeo: Transformer is all you need for cross-view image geo-localization, 2022

    Sijie Zhu, Mubarak Shah, and Chen Chen. TransGeo: Transformer is all you need for cross-view image geo-localization, 2022. URLhttps://arxiv.org/abs/2204.00097. CVPR 2022

    Pith/arXiv arXiv 2022

  5. [5]

    LoFTR: Detector-free local feature matching with transformers, 2021

    Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. LoFTR: Detector-free local feature matching with transformers, 2021. URLhttps://arxiv.org/abs/2104.00680

    Pith/arXiv arXiv 2021

  6. [6]

    RoMa: Robust dense feature matching, 2023

    Johan Edstedt, Qiyu Sun, Georg Bökman, Mårten Wadenbäck, and Michael Felsberg. RoMa: Robust dense feature matching, 2023. URLhttps://arxiv.org/abs/2305.15404. CVPR 2024

    Pith/arXiv arXiv 2023

  7. [7]

    DUSt3R: Geometric 3d vision made easy, 2024

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. DUSt3R: Geometric 3d vision made easy, 2024. URLhttps://arxiv.org/abs/2312.14132

    Pith/arXiv arXiv 2024

  8. [8]

    Grounding image matching in 3d with MASt3R, 2024

    Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with MASt3R, 2024. URL https://arxiv.org/abs/2406.09756

    Pith/arXiv arXiv 2024

  9. [9]

    VGGT: Visual geometry grounded transformer, 2025

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. VGGT: Visual geometry grounded transformer, 2025. URLhttps://arxiv.org/abs/2503.11651. CVPR 2025

    Pith/arXiv arXiv 2025

  10. [10]

    Reloc3r: Large-scale training of relative camera pose regression for generalizable, fast, and accurate visual localization,

    Siyan Dong, Shuzhe Wang, Shaohui Liu, Lulu Cai, Qingnan Fan, Juho Kannala, and Yanchao Yang. Reloc3r: Large-scale training of relative camera pose regression for generalizable, fast, and accurate visual localization,

  11. [11]

    University-1652: A multi-view multi-source benchmark for drone-based geo-localization

    Zhedong Zheng, Yunchao Wei, and Yi Yang. University-1652: A multi-view multi-source benchmark for drone-based geo-localization. InProceedings of the 28th ACM International Conference on Multimedia, pages 1395–1403, 2020. doi: 10.1145/3394171.3413896. URLhttps://doi.org/10.1145/3394171.3413896

    arXiv 2020

  12. [12]

    Multiple-environment self-adaptive network for aerial-view geo-localization.Pattern Recognition, 152:110363, 2024

    Tingyu Wang, Zhedong Zheng, Yaoqi Sun, Chenggang Yan, Yi Yang, and Tat-Seng Chua. Multiple-environment self-adaptive network for aerial-view geo-localization.Pattern Recognition, 152:110363, 2024. doi: 10.1016/j. patcog.2024.110363. URLhttps://doi.org/10.1016/j.patcog.2024.110363

    arXiv 2024

  13. [13]

    Towards Natural Language-Guided Drones: GeoText-1652 Benchmark with Spatial Relation Matching

    Meng Chu, Zhedong Zheng, Wei Ji, Tingyu Wang, and Tat-Seng Chua. Towards Natural Language-Guided Drones: GeoText-1652 Benchmark with Spatial Relation Matching. InComputer Vision – ECCV 2024, pages 213–231, 2024. doi: 10.1007/978-3-031-73247-8_13. URLhttps://arxiv.org/abs/2311.12751

    Pith/arXiv arXiv 2024

  14. [14]

    Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, 2018

    Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics, 2018. URLhttps://arxiv.org/abs/1705.07115. CVPR 2018

    Pith/arXiv arXiv 2018

  15. [15]

    GradNorm: Gradient normalization for adaptive loss balancing in deep multitask networks, 2018

    Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. GradNorm: Gradient normalization for adaptive loss balancing in deep multitask networks, 2018. URL https://arxiv.org/abs/1711.02257. ICML 2018

    Pith/arXiv arXiv 2018

  16. [16]

    Gradient surgery for multi-task learning, 2020

    Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning, 2020. URLhttps://arxiv.org/abs/2001.06782. NeurIPS 2020. 7

    Pith/arXiv arXiv 2020

  17. [2025]

    CVPR 2025

    URLhttps://arxiv.org/abs/2412.08376. CVPR 2025

    Pith/arXiv arXiv 2025