{"id":"610e353a-bb31-4215-9ea8-f1430aefa29c","arxiv_id":"2505.01207","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A plug-and-play auxiliary branch that regresses pairwise camera translations improves sparse-view camera pose estimation across RelPose++ and Forge-2D on CO3D and IMC PhotoTourism.","lead":"This paper adds a small training-time side module, T-Graph, to existing camera pose estimators. The module predicts relative translations between each pair of cameras, and the extra supervision improves sparse-view pose accuracy by a few percent on two benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2's RelPose++/CO3D comparison uses published baselines rather than retrained ones, so the headline 1–6% camera-center improvement may reflect training-condition differences. This is the most load-bearing concern; retraining the baseline under the authors' exact pipeline would settle it.","rationale":"The paper proposes a training-only auxiliary loss that supervises pairwise translations, which is a simple and plausible mechanism for improving sparse-view pose estimation. The empirical claim is supported by four experiment groups. The strongest evidence would be controlled comparisons where baseline and augmented models are trained with identical settings; EG2–EG4 provide this, and they show positive camera-center improvements. However, the flagship EG1 result for RelPose++ on CO3D (Table 2) is the only one not using a retrained baseline, and it is prominently cited in the abstract's '1% to 6%' improvement. Because the authors used a different GPU and batch size for EG1 (H100, batch 22) than for EG2–EG4 (A100, batch 8/16/16), and because the published RelPose++ numbers were produced elsewhere, the Table 2 margins could reflect optimization or hyperparameter differences rather than T-Graph. This is the load-bearing assumption in the central claim. The paper would be substantially strengthened by retraining the baseline under the authors' exact pipeline and reporting error bars. Given the controlled experiments in Tables 3–5, the concern is addressable and does not warrant rejection; it does warrant a conditional verdict, consistent with the reader's assessment. We therefore recommend no change to the verdict.","tokens_in":14308,"tokens_out":9484,"duration_ms":78844,"concrete_test":"Retrain the RelPose++ baseline without T-Graph using the authors' EG1 pipeline: same codebase, CO3D 41-category training split, H100 GPU, batch size 22, AdamW, learning rate 1e-5, same sequence order file, and five random sampling trials per viewpoint count (2–8). Compare the resulting camera-center, rotation, and translation accuracies against the published RelPose++ numbers in Table 2 and against Ours(pair-t) and Ours(relative-t). If the retrained baseline differs from the published values by more than ~0.5 percentage points (the typical gain in Tables 3–5), or if the gap to Ours(pair-t) shrinks, the comparison is unreliable. Repeat with 3–5 seeds to bound variance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 explicitly states: 'the evaluation results of RelPose++ on CO3D are directly taken from its original publication [15], and thus, no additional training was conducted in our experiments.' The authors then trained their T-Graph-augmented RelPose++ on a single H100 with batch size 22, 'keeping all other hyperparameters identical to those used in RelPose++.' This is not a controlled comparison: the published RelPose++ numbers may come from a different GPU, batch size, number of epochs, learning-rate schedule, or random seed. Since the central claim of 'camera center accuracy improving by 1% to 6%' (abstract) and the 'consistent improvements' (Sec. 4.2) are anchored by Table 2, any systematic difference in training conditions could inflate or deflate the reported gains. The remaining tables (EG2–EG4) do train baselines under identical settings, and their camera-center improvements are positive but often small (0.2–2 percentage points), with no error bars or multiple seeds reported. Thus the most load-bearing assumption is that the published RelPose++/CO3D baselines are equivalent to what would be obtained under the authors' training pipeline. If that assumption fails, the headline improvement for the primary baseline is not attributable to T-Graph.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes T-Graph, a lightweight, plug-and-play auxiliary branch for sparse-view camera pose estimation. During training, T-Graph takes pairs of image features, predicts pairwise translations through a shared MLP, and adds an L1 loss under either a relative-t or a pair-t representation; the branch is removed at inference, so the baseline architecture is unchanged. The method is integrated into RelPose++ and a 2D-branch variant of Forge and evaluated on CO3D and IMC PhotoTourism for 2 to 8 viewpoints. The authors report consistent improvements across the four experiment groups, with camera-center accuracy improving by 1% to 6%, and conclude that pairwise translation supervision is a cost-effective way to improve pose estimation.","tokens_in":14586,"tokens_out":7098,"duration_ms":68917,"significance":"If the claimed gains hold, T-Graph would be a useful training-time regularizer for sparse-view pose estimation: it adds only a small number of parameters, does not change inference cost, and is architecturally agnostic. The paper also proposes two pairwise translation representations with a plausible geometric interpretation, which is a useful conceptual contribution for choosing supervision under different camera configurations. The empirical evaluation covers two public datasets and two representative baseline families, and the qualitative visualizations support the general direction. However, the strongest quantitative claim is not yet rigorously established because the headline CO3D/RelPose++ comparison relies on published numbers rather than retrained baselines, error bars are absent, and the added MLP capacity is not ablated. The paper is clearly written and the idea is simple and easy to reproduce, but the evidence as presented supports a moderate, not a strong, claim of consistent improvement.","major_comments":[{"comment":"The comparison between RelPose++ and T-Graph on CO3D is not a controlled experiment. The manuscript states that the RelPose++ evaluation results are taken directly from the original publication and that no additional training was conducted, while the T-Graph variants are trained on a single H100 with batch size 22. Different hardware, batch size, effective epoch count, learning-rate schedule, or random seed in the published baseline could account for part or all of the reported gains. The authors should retrain RelPose++ under their exact pipeline and report those numbers, or otherwise the 1–6% headline improvement in camera-center accuracy should be substantially qualified.","section":"Sec. 4.1, Table 2"},{"comment":"The results are reported without error bars or multiple seeds; for EG 2–4 the paper explicitly uses a fixed random seed and a single random sampling trial per viewpoint count. Many of the reported improvements are small (typically 0.2–2 percentage points), and some entries regress relative to the baseline, for example Table 3 translation accuracy at 2 views (0.589 vs. 0.595) and Table 2 rotation accuracy at 2 views for relative-t (0.684 vs. 0.698). The claim of 'consistent improvements' in Sec. 4.2 and the abstract therefore goes beyond what the reported numbers establish. The authors should provide variance across seeds and sampling trials, or temper the claim accordingly.","section":"Sec. 4.1, Tables 2–5"},{"comment":"There is no ablation that controls for the added capacity of the 6-layer MLP branch. Table 6 shows that T-Graph increases model size by 5–11%, and since the branch is active only during training, the observed gains could partly come from extra model capacity, the additional training signal in general, or regularization effects. An ablation that replaces the pairwise-translation loss with an auxiliary loss of matched capacity (for example, the same MLP predicting scalar targets or per-camera absolute translations) would be needed to attribute the improvement specifically to pairwise translation supervision.","section":"Sec. 3.1, Table 6"},{"comment":"The pair-t formulation is geometrically inaccurate for the general case used in the experiments. The paper defines W_k as the point that minimizes the distance to both optical axes, but then asserts that its coordinates in the two camera frames are (0,0,D_i) and (0,0,D_j). When the optical axes do not exactly intersect, which is the generic case and is explicitly contemplated in the 'approximately intersect' discussion in Sec. 5, W_k is not on either optical axis, so the vector from a camera center to W_k has off-axis components and is not simply (0,0,D). The rotation-disentanglement claim should be restated, or the representation should be redefined (e.g., using the foot of the perpendicular from W_k onto each axis) so that the stated coordinates are correct.","section":"Sec. 3.2, Fig. 3(b)"}],"minor_comments":[{"comment":"The dataset name is written as 'C03D' in the abstract and in Sec. 4.1; this should be 'CO3D' throughout.","section":"Abstract, Sec. 4.1"},{"comment":"The paper alternates between 'T-Graph' and 'T-graph' (for example, the abstract uses both). Please use a single consistent name.","section":"Throughout"},{"comment":"The notation in Eq. (1) is not fully explained: X^w is called 'the point in the world frame' and X^c_j 'the corresponding point in the camera frame of C_j', but the relation between these points and the translation vector t_{i→j} should be stated more explicitly, including the coordinate-frame definitions for R_{i→j}.","section":"Eq. (1)"},{"comment":"The description of the Forge-2D integration is somewhat hard to follow: the text says self-attention is applied to the first image feature so that all k pose features can serve as inputs to T-Graph. A short pseudo-code or tensor-shape description would clarify how the k pose features are obtained and used.","section":"Sec. 4.1"},{"comment":"The statement that 'the camera center accuracy at a threshold of 0.2 is always 1 when the number of views is 2' is given without justification. A one-sentence explanation (for example, that two camera centers can always be aligned by a similarity transform) would help the reader.","section":"Sec. 4.2"}],"recommendation":"major_revision","confidential_remarks":"The paper's remote-sensing framing is mostly motivational; the experiments are standard computer-vision benchmarks on CO3D and IMC PhotoTourism, which may still be within the journal's scope but is worth noting. The main issue is that the headline result rests on Table 2, which is not a controlled comparison, and the empirical claims lack error bars. I would recommend requesting a retrained RelPose++ baseline under the authors' pipeline, multiple seeds, and an ablation for the added MLP capacity. The pair-t geometry issue in Sec. 3.2 should also be corrected or the claims weakened before the paper is considered further."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean, incremental engineering paper with a useful auxiliary-supervision idea and an honest discussion of when each representation helps. The central idea—add a fully connected graph of pairwise translation targets, computed from ground truth, as an extra training loss—is simple and plausible. The pair-t parameterization (putting the origin at the optical-axis intersection) is the most original piece, and the authors are transparent that it suits center-facing camera configurations while relative-t suits roughly parallel ones. The experiments across two baselines and two datasets, with the two representations behaving as predicted, is real evidence. EG2–EG4 retrain baselines under the same protocol, and gains are small but mostly positive.\n\nSoft spots, in order:\n\n1. The headline comparison (EG1, RelPose++ on CO3D) uses published numbers instead of retraining the baseline. The authors admit this in Sec. 4.1. Since batch size, scheduler, seeds, and hardware could differ, the 1–6% camera-center gain in the abstract is not a controlled measurement. The stress-test note is correct, and this is the load-bearing caveat.\n\n2. No error bars or multiple seeds anywhere. The paper reports five sampling trials for EG1, but that is test-time variation, not training stability. Given the gains are often a couple of percentage points, we do not know whether they are within noise.\n\n3. No ablation controls for added MLP capacity. The T-Graph branch shares features and adds a 6-layer MLP; one wants a baseline with an equivalent extra head or a different auxiliary loss to rule out capacity or optimization effects.\n\n4. Minor: DUSt3R and other pairwise methods are not discussed, which weakens the novelty framing. And the text says \"consistent improvements\" but several individual cells regress slightly (e.g., relative-t rotation at 2 views in Table 2, translation at 2 views in Table 5). That is not fatal; the trend is real.\n\nThe math checks out, there are no hidden fitted constants, and the public datasets make the work reproducible. Citation pattern is fine; self-citation is not an issue.\n\nWho is this for? People working on sparse-view pose estimation and lightweight training-time add-ons. It deserves a serious referee, but a reviewer should ask for retrained RelPose++/CO3D baselines and error bars before accepting the headline claim. My own verdict: conditional accept with those revisions.","headline":"A lightweight training-only pairwise translation module that seems to help sparse-view pose estimation, but the main RelPose++/CO3D comparison relies on borrowed baselines and no error bars, so treat the headline gain as plausible rather than proven.","tokens_in":15085,"tokens_out":1937,"would_cite":true,"duration_ms":19035,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Pairwise translation supervision, organized as a fully connected graph, improves sparse-view camera pose estimation.","keywords":["camera pose estimation","sparse-view","pairwise translation","translation graph","relative-t","pair-t","auxiliary supervision","plug-and-play module"],"falsifier":"Retrain RelPose++ from scratch on CO3D with the exact hyperparameters, batch sizes, and random seeds used for the T-Graph runs, then compare camera center accuracy at 3 to 8 views; if the gap between baseline and T-Graph disappears or falls below 1%, the claimed improvement is at least partly an artifact of training conditions.","tokens_in":14142,"feed_emoji":"📷","tokens_out":7050,"duration_ms":60721,"temperature":0.7,"pith_summary":"This paper claims that sparse-view camera pose estimation improves when the model is also supervised on the translation between every pair of cameras, not just each camera's translation to a fixed world origin. The proposed T-Graph is a lightweight module that shares the baseline feature extractor, predicts pairwise translations through a small MLP, and arranges them as a fully connected graph with cameras as nodes. It is active only during training and removed at inference, so any gains come without slowing the deployed model. Across the RelPose++ and Forge-2D baselines on the CO3D and IMC PhotoTourism datasets, the authors report consistent gains, with camera center accuracy improving by 1% to 6% from 2 to 8 viewpoints.","feed_headline":"Pairwise translation graph lifts sparse-view camera pose accuracy","feed_subtitle":"A training-only add-on improves localization on two methods and two datasets, adding no inference cost.","key_machinery":"The load-bearing mechanism is the fully connected translation graph: for $n$ cameras there are $\\binom{n}{2}$ edges, and each edge is produced by the same lightweight MLP regressor from the pair of image features $(f_i, f_j)$. The graph replaces sparse per-camera translation regression with dense pairwise supervision, and the scaling factors $k_1 = n / (2 \\binom{n}{2})$ for pair-t and $k_2 = n / \\binom{n}{2}$ for relative-t keep the auxiliary loss comparable to the baseline translation loss. The two representations matter because pair-t removes the rotation-translation coupling by putting the origin at the optical-axis intersection, while relative-t keeps a simple local frame that remains stable when optical axes are nearly parallel. Because the module is dropped at inference, it can only improve accuracy by shaping the shared feature extractor during training.","core_discovery":"The paper's central claim is that pairwise translation information is a valuable and underused supervision signal for sparse-view pose estimation. T-Graph builds a complete graph whose nodes are cameras and whose edges are translations between camera pairs, predicted by a shared MLP from paired image features. These pairwise predictions are not used as final output; they contribute an L1 loss during training that steers the shared feature extractor toward globally informative representations. Two edge encodings are introduced: relative-t, which sets the world origin at one camera and remains entangled with rotation, and pair-t, which sets the origin at the intersection of the two optical axes and thereby disentangles translation from rotation. Empirically, pair-t suits center-facing camera configurations such as CO3D, while relative-t suits roughly parallel configurations such as IMC PhotoTourism.","pith_inferences":["The same pairwise-supervision idea could be applied to relative rotations, producing a fully connected rotation graph that may complement the translation graph and improve pose consistency further.","A learned combination or soft switch between relative-t and pair-t could remove the need to choose the representation manually for each dataset.","Since the module is removed at inference, any benefit must be mediated through the shared feature extractor; this makes T-Graph a useful probe for whether explicit geometric supervision during training produces more pose-aware features, a claim that could be checked by probing feature similarity."],"forward_implications":["Camera center accuracy improves by 1% to 6% across 2 to 8 viewpoints on both CO3D and IMC PhotoTourism when T-Graph is added.","The gains transfer across two architecturally distinct baselines, RelPose++ (energy-based rotation model) and Forge-2D (purely discriminative), which supports the module's generalizability.","The choice of representation should follow the geometry of the camera distribution: pair-t for center-facing cameras, relative-t for mostly parallel cameras.","Because T-Graph is removed at inference, it adds no runtime cost and only modest parameter growth (+5% for RelPose++, +11% for Forge-2D).","Pairwise ground-truth translations are a cost-effective additional training signal for sparse-view pose estimation."],"supporting_citations":[{"why":"RelPose++ is the primary baseline; T-Graph is plugged into it and the CO3D evaluation protocol and metrics are inherited from it.","marker":"[15]"},{"why":"Forge is the second baseline, and its 2D branch is used as Forge-2D to test T-Graph on a purely discriminative pose model.","marker":"[19]"},{"why":"CO3D provides the object-centric dataset whose center-facing camera distribution motivates the pair-t representation.","marker":"[20]"},{"why":"IMC PhotoTourism provides the landmark dataset with mostly parallel cameras that motivates the relative-t representation.","marker":"[21]"},{"why":"RelPose supplies the fixed-seed sampling protocol and the rotation-focused energy-based formulation that RelPose++ builds on.","marker":"[34]"},{"why":"COLMAP supplies the ground-truth camera poses used for both datasets and represents the classical SfM approach the work contrasts with.","marker":"[12]"},{"why":"The least-squares similarity transform alignment underlies the camera center and translation accuracy metrics.","marker":"[36]"}],"fun_headline_variants":["Training-only translation graph refines sparse-view camera poses","Graph of pairwise translations improves pose from few viewpoints","Translation relationships between cameras sharpen sparse-view pose","Sparse-view camera localization gains from pairwise translation model","Pairwise translation loss enhances pose estimation in sparse views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison on CO3D assumes the published RelPose++ numbers were obtained under the same training conditions as the authors' runs, since those baselines were not retrained; if the training settings differ, the reported improvements could be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Training-only translation graph refines sparse-view camera poses","Graph of pairwise translations improves pose from few viewpoints","Translation relationships between cameras sharpen sparse-view pose","Sparse-view camera localization gains from pairwise translation model","Pairwise translation loss enhances pose estimation in sparse views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000854,"raw_usage":{"total_tokens":3728,"prompt_tokens":981,"completion_tokens":2747,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":2674}},"tokens_in":597,"tokens_out":2747,"duration_ms":21386,"temperature":1.0,"reasoning_tokens":2674,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:23:38.384947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain RelPose++ from scratch on CO3D with the exact hyperparameters, batch sizes, and random seeds used for the T-Graph runs, then compare camera center accuracy at 3 to 8 views; if the gap between baseline and T-Graph disappears or falls below 1%, the claimed improvement is at least partly an artifact of training conditions.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RelPose++ is the primary baseline; T-Graph is plugged into it and the CO3D evaluation protocol and metrics are inherited from it."},{"cited_title":"Jiang, Z","cited_arxiv_id":null,"evidence_quote":"Forge is the second baseline, and its 2D branch is used as Forge-2D to test T-Graph on a purely discriminative pose model."},{"cited_title":"Reizenstein, R","cited_arxiv_id":null,"evidence_quote":"CO3D provides the object-centric dataset whose center-facing camera distribution motivates the pair-t representation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"IMC PhotoTourism provides the landmark dataset with mostly parallel cameras that motivates the relative-t representation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RelPose supplies the fixed-seed sampling protocol and the rotation-focused energy-based formulation that RelPose++ builds on."},{"cited_title":"Umeyama, Least-squares estimation of transformation parameters be- tween two point patterns, IEEE Transactions on Pattern Analysis & Machine Intelligence 13 (04) (1991) 376–380","cited_arxiv_id":null,"evidence_quote":"The least-squares similarity transform alignment underlies the camera center and translation accuracy metrics."}],"review_version":1}