{"id":"5c5c8528-5fbf-4aa1-8ca6-a856786f4ec1","arxiv_id":"2607.02711","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"A 10.7k-parameter residual PE/RoPE adapter, fit online on a short fisheye clip, recovers 2–12× better rotation from frozen pinhole 3D foundation models while keeping single-pass inference.","lead":"RayTun3R fixes fisheye failures in frozen 3D foundation models by adapting only tiny positional-encoding residuals online. It cuts pose error sharply without multi-view inference cost or large fine-tunes, which matters for robots and wide-FOV cameras.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged supervision caveat.","rationale":"The paper's argument is: pinhole PE bias (constant Jacobian) is a primary failure mode under fisheye; correcting only PE/RoPE residuals (plus parameter-free grid/tokenization fixes) with self-supervised geometric losses recovers pose while freezing the backbone. Evidence is consistent and multi-faceted. The only material risk is that the geometric losses could be optimizing toward matcher-specific artifacts rather than pure camera geometry; the authors already mitigate this by freezing the MAGSAC++ target and showing AnyCalib robustness. That risk is real but ordinary for online geometric adaptation and does not justify changing the ACCEPT verdict. Code release and error bars would further strengthen confidence but are not required to accept the stated claim.","tokens_in":19024,"tokens_out":453,"duration_ms":5240,"concrete_test":"On one ETH3D and one FIORD sequence, re-fit RayTun3R using only the L_reproj + smoothness + PE regularizers (drop L_pose entirely) and, separately, replace UFM matches with synthetic ground-truth correspondences from the provided poses/depth where available; if rotation/translation errors rise by more than ~2× relative to the full objective, the external-matcher premise is load-bearing; otherwise the claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is well supported by the PE-Jacobian diagnosis (Fig. 2), multi-backbone multi-dataset pose gains (Tables 1–2), component ablations isolating learned PE residuals (Table 4a, Sec. G), and competitive depth (Table 3). The reader's weakest assumption—that UFM matches plus once-computed MAGSAC++ poses form sufficiently accurate, camera-geometry-independent supervision for a short 30-window fit—is the genuine soft spot, but it is already correctly identified and does not overturn the empirical pattern: gains hold under AnyCalib calibration (Table 5), across FOVs 110°–200°, and when parameter-free corrections alone fail. No stronger internal inconsistency or hidden assumption appears load-bearing.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper diagnoses a pinhole bias in the positional encodings of recent 3D foundation models (DA3, VGGT, π³) via the local Jacobian of absolute PE (Fig. 2, Sec. 3), and proposes RayTun3R: a frozen-backbone adapter that learns small residual radial/angular lookup tables for absolute PE and RoPE (Eqs. 5–6), plus parameter-free patch undistortion, border-token handling, and DPT-grid coordinate correction. The adapter (10,752 parameters on DA3-Small) is fit online on ~30 three-frame windows with geometric losses that use fixed external UFM matches and a once-computed MAGSAC++ pose target (Sec. 4.3, Eqs. 8–13). Across five fisheye datasets (110°–200° FOV) and multiple backbones, it reduces rotation error by roughly 2–12× vs. the unadapted model, outperforms LoRA and CalTok with far fewer parameters, improves pose over Center-PH/Multi-PH without multi-view cost, and remains competitive on depth (Tables 1–3, 6).","tokens_in":19259,"tokens_out":977,"duration_ms":9398,"significance":"If the results hold, the work offers a practical, low-cost route to reuse pinhole-trained 3D foundation models on fisheye imagery without retraining or multi-view projection. Strengths include a clear geometric diagnosis (PE Jacobian), extreme parameter efficiency, multi-backbone multi-dataset evaluation, component ablations isolating learned PE residuals (Table 4a, Sec. G), and sensitivity checks under predicted calibration (Table 5). The approach is falsifiable and immediately useful for robotics and mapping pipelines that already deploy these models. Code release is promised, which would further strengthen reproducibility.","major_comments":[{"comment":"The supervision pipeline in Sec. 4.3 (UFM matches + once-computed MAGSAC++ pose target, Eqs. 8–9) is the load-bearing soft spot. Gains are large and consistent, but the paper does not quantify how much residual matcher bias under strong fisheye distortion remains after the fixed-target design. A short controlled experiment—e.g., synthetic fisheye with known ground-truth matches, or deliberate corruption of UFM confidence—would make the isolation of PE bias more airtight. This does not overturn the empirical pattern, but it is the main correctness-risk concern for the central claim.","section":null},{"comment":"Evaluation protocol (Sec. 5): adaptation uses 30 filtered windows (optical-flow ≥2 px) from the same sequence that is later evaluated. While this matches the intended online setting, there is no multi-seed variance, no held-out sequence transfer within a camera, and no error bars. For a journal claim of 2–12× rotation reduction, at least seed-level variance or a leave-one-sequence-out check on one multi-sequence dataset (e.g., KITTI-360 or TUM-VI) would strengthen confidence that the adapter is not over-fitting sequence-specific appearance.","section":null}],"minor_comments":[{"comment":"Table 4a: the full RayTun3R row is not uniformly best on every metric (some ablations have slightly lower R° or t°). A one-sentence clarification that the full model is selected for lowest dreproj / overall balance would avoid reader confusion.","section":null},{"comment":"Fig. 2 caption and panels: the switch from “pretrained σ1 / det” to “adapted” is clear, but the analytical fisheye reference curves are not overlaid on the same axes; adding them would make the visual match more immediate.","section":null},{"comment":"Notation: κ and κ⁻¹ are introduced in Sec. 3, but the precise fisheye model (KB vs. EUCM) used per dataset is only mentioned in passing; a short table or appendix note would help reproducibility.","section":null},{"comment":"Related work: contemporaneous Fisheye3R and FishRoPE are cited; a slightly sharper sentence on the frozen-backbone / self-supervised distinction would further clarify novelty.","section":null},{"comment":"Typos / polish: “π3” vs. “π³” inconsistency; “drive 0000, cam02” style labels in Table 4a could be standardized with the main tables.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is solid and well above the bar for a solid CV journal contribution. The only reason I did not recommend outright accept is the missing variance / matcher-sensitivity check; both are straightforward to address in revision. Scope and novelty fit well; no citation or disclosure concerns."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful bit is simple: they show the pretrained positional embeddings of DA3/VGGT/π³ carry a pinhole Jacobian (flat singular values and det vs radius in Fig. 2), then fix it with ~10.7k residual radial/angular PE + RoPE tables plus a few parameter-free camera-aware patches. That is the real contribution, not another generic PEFT layer.\n\nWhat they do well is keep the backbone frozen, train online on 30 three-frame windows with fixed UFM matches and a once-computed MAGSAC pose target, and still get 2–12× lower rotation error across five real fisheye sets (110–200°) and three backbone families. Tables 1–2 and the component ablations (Table 4a, Sec. G) isolate the learned PE residual as the main driver; LoRA and CalTok lose while using far more parameters; Multi-PH costs ~4× at inference. Depth stays competitive with Center-PH while keeping the full FOV. AnyCalib sensitivity (Table 5) and the loss/bin ablations are honest. Citations cover the contemporaneous Fisheye3R/FishRoPE work and correctly distinguish the frozen self-supervised setting.\n\nSoft spots are ordinary engineering ones, not structural. Supervision rides on external UFM + MAGSAC; if the matcher is systematically wrong under extreme distortion or motion is near-static, the geometric losses weaken. No error bars or multi-seed runs. Adaptation is camera-specific and assumes mostly radial distortion. Code is promised, not shipped. None of that overturns the pattern in the tables.\n\nThis is for people who already run feed-forward 3D models on robots or vehicles with wide-FOV sensors and want a cheap online fix rather than a new architecture. The math is light but the diagnosis and the empirical isolation are solid. I would send it to peer review; it deserves a serious referee, not a desk reject.","headline":"Clean, targeted PE residual adapter that actually makes frozen 3D foundation models usable on real fisheye without multi-view cost.","tokens_in":19890,"tokens_out":492,"would_cite":true,"duration_ms":5345,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Pinhole bias in positional encodings is why 3D foundation models fail on fisheye images, and a tiny residual adapter can fix it online.","keywords":["3D foundation models","fisheye cameras","positional encodings","online adaptation","camera geometry","depth and pose estimation","parameter-efficient fine-tuning"],"falsifier":"On a high-FOV fisheye sequence with accurate ground-truth poses, replace the external matcher with perfect correspondences (or deliberately corrupt them) and check whether the PE residual still recovers the claimed rotation and translation gains; if the gains vanish under clean matches or survive under strongly biased ones, the supervision claim fails.","tokens_in":19928,"feed_emoji":"📷","tokens_out":680,"duration_ms":5701,"temperature":0.7,"pith_summary":"Modern 3D foundation models recover depth and pose from a few images, but they collapse when the camera is a fisheye rather than a pinhole. The paper argues that a large part of this failure lives in the positional encodings: those encodings were learned under a constant local geometry that only a pinhole camera has, so the same image-grid step no longer means the same change of viewing direction under fisheye optics. RayTun3R freezes the backbone and corrects only the camera-dependent pieces—small residual lookup tables on absolute and rotary positional encodings, plus parameter-free fixes to patch tokenization and prediction-grid coordinates. The adapter has about eleven thousand parameters, is fit on a short self-supervised temporal segment, and then runs at the original single-pass cost on the rest of the sequence. Across outdoor, handheld, and indoor fisheye data from 110° to 200° field of view, and across several frozen backbones, it cuts rotation error by factors of two to twelve relative to the unadapted model, beats generic low-rank fine-tuning with far fewer parameters, and improves pose over multi-view cropping baselines without their extra inference cost while staying competitive on depth.","feed_headline":"Tiny PE adapter fixes fisheye failure in 3D foundation models","feed_subtitle":"About 11k parameters cut rotation error 2–12× on 110–200° cameras without retraining the backbone","key_machinery":"RayTun3R: zero-initialized residual radial/angular lookup-table corrections to absolute and rotary positional encodings, trained online with geometric losses on a short multi-frame segment while the foundation model stays frozen.","core_discovery":"The local Jacobian of pretrained positional embeddings is nearly radius-independent and therefore encodes a pinhole spatial prior; once that prior is corrected by lightweight residual PE/RoPE adapters (plus parameter-free tokenization and grid fixes), frozen 3D foundation models recover coherent fisheye depth and pose without retraining the backbone.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["11k-param PE residual fixes fisheye bias in frozen 3D models","RayTun3R corrects pinhole PE prior for online fisheye adaptation","Lightweight residual adapters recover fisheye pose without backbone train","PE and RoPE residuals cut fisheye rotation error 2-12x at 11k params","Parameter-free tokenization plus PE fixes enable fisheye 3D recovery"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that external matches and a once-computed relative pose from a short, non-static clip give clean enough geometric supervision to isolate the camera bias rather than scene-specific or matcher-specific errors.","fun_headline_variants_meta":{"raw":{"variants":["11k-param PE residual fixes fisheye bias in frozen 3D models","RayTun3R corrects pinhole PE prior for online fisheye adaptation","Lightweight residual adapters recover fisheye pose without backbone train","PE and RoPE residuals cut fisheye rotation error 2-12x at 11k params","Parameter-free tokenization plus PE fixes enable fisheye 3D recovery"]},"model":"grok-4.5","effort":"low","cost_usd":0.004884,"raw_usage":{"total_tokens":1414,"prompt_tokens":804,"num_sources_used":0,"completion_tokens":89,"cost_in_usd_ticks":48840000,"prompt_tokens_details":{"text_tokens":804,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":521,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":804,"tokens_out":89,"duration_ms":4841,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T07:33:38.106123+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a high-FOV fisheye sequence with accurate ground-truth poses, replace the external matcher with perfect correspondences (or deliberately corrupt them) and check whether the PE residual still recovers the claimed rotation and translation gains; if the gains vanish under clean matches or survive under strongly biased ones, the supervision claim fails.","supporting_citations":[],"review_version":1}