acceptodds
Under review as a conference paper at ICLR 2027

FFBA: Feed-Forward Bundle Adjustment

Abstract

We propose feed-forward bundle adjustment (FFBA), which recasts per-scene iterative BA as a single forward pass of a learned refinement model with update rules shared across scenes. Given initial cameras and geometry from a geometric foundation model (GFM) and fixed multi-view tracks, FFBA constructs a sparse heterogeneous graph of cameras, 3D points and 2D observations. It then refines poses, focal lengths and 3D points through a geometry-conditioned graph network stacked with geometric graph aggregation layers. These layers update observation, camera and point features via graph message propagation, and filter unreliable observations through uncertainty-guided confidence estimation. A confidence-weighted reprojection loss trains the whole framework end-to-end using only in-house data, without manual annotation. To support training across datasets and GFMs, we develop a standardized sample-construction pipeline that yields over 0.9M samples from seven datasets and three GFMs with shared multi-view tracks. Experiments on three in-domain and four cross-domain datasets show that FFBA consistently improves reprojection consistency and camera accuracy over GFM initializations. It also remains competitive with classical BA and EPO under domain shift, and reduces optimization cost by more than an order of magnitude, translating into a clear end-to-end efficiency advantage that is expected to grow with scene scale. Code and dataset will be shared upon paper acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Bundle Adjustment, Feed-Forward, 3D Reconstruction, Graph Attention Networks

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_CQ21OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Feed-Forward Bundle Adjustment (FFBA), which replaces per-scene iterative optimization with a single forward pass through a learned refinement network. Given initial camera parameters and 3D points from a geometric foundation model, together with fixed multi-view correspondences, FFBA constructs a heterogeneous graph comprising camera, 3D point, observation, and global nodes. Eight geometric graph aggregation layers jointly refine camera poses, focal lengths, and 3D coordinates. The framework is trained without manual annotations using a confidence-weighted reprojection loss. The authors construct approximately 0.9 million samples using three foundation models. Experiments on in-domain and cross-domain datasets demonstrate improvements over the initial geometry and substantially reduced refinement time, while ablation studies assess the contributions of the key components.

Critical strengths
  1. To my knowledge, this is the first work I have encountered that explores a feed-forward approach to bundle adjustment. The method design and experimental results demonstrate the feasibility of replacing per-scene iterative optimization with learned feed-forward refinement. This provides a promising alternative to conventional BA and is a substantive contribution supporting acceptance.

  2. bottlenThe paper addresses a meaningfuleck in pipelines built on geometric foundation models: their predictions retain geometric errors, while iterative refinement through conventional or differentiable BA can erode their speed advantage. FFBA substantially reduces refinement time while maintaining competitive accuracy. Although the results do not show a dramatic leap in accuracy, the approach offers an insightful direction for future research on efficient geometric refinement.

Critical weaknesses

The cross-domain experiments demonstrate FFBA’s ability to improve the initial geometry estimates, but its accuracy advantage over conventional BA is not consistent. FFBA achieves lower relative pose AUC than BA on three of the four cross-domain datasets, suggesting limitations in the transferability of the learned update rules across data distributions. Further analysis of the scene conditions associated with these accuracy gaps would help clarify when FFBA is a suitable alternative to conventional BA.

Questions for authors

FFBA improves the initial estimates across all evaluated cross-domain datasets, but its camera accuracy does not consistently surpass that of BA and EPO. This raises questions about the applicability of the learned update rules: to what extent does performance depend on the scale and distribution of the training data, as well as the GFMs used during training? Would FFBA require retraining or fine-tuning if the upstream GFM architecture or its error distribution changed? Furthermore, as GFMs increasingly incorporate geometric consistency constraints and potentially optimization mechanisms, how do the authors envision FFBA’s role as a standalone refinement module or as an integral component of a geometric foundation model?

2 comments

We thank the reviewer for recommending acceptance and recognizing the value of FFBA. Below, we discuss the dependence on training data, present results on unseen GFMs, and outline future integration with GFMs.


1. When does FFBA fall short of BA?

We agree that FFBA does not consistently match BA's camera accuracy: BA achieves higher relative-pose AUC on ETH3D, LLFF, and Mip360, whereas FFBA performs better on T&T (Tab. 2). Our experience suggests that difficult initializations and weak geometric constraints, such as limited parallax or sparse view overlap, can challenge refinement with a fixed update budget; iterative BA can instead continue adapting to each scene.


2. Dependence on training data and initialization sources

FFBA learns geometric corrections from data, so its performance depends on how well training covers the scene structures and initialization errors encountered at test time. We train on three datasets and three GFMs with shared underlying tracks to broaden this coverage (Sec. 4). Our experience motivates balancing initialization quality and error patterns, rather than increasing sample count alone. We have not isolated the effects of training scale, data distribution, and initializer diversity, so we cannot quantify their individual contributions. Nevertheless, the unseen-GFM experiments below show that useful transfer is possible beyond the training initializers.


3. Does changing the upstream GFM require retraining?

Thank you for suggesting this evaluation, which directly tests transfer to unseen initializers. We evaluated MapAnything and DA3-Large, neither used during FFBA training, on three in-domain and four cross-domain datasets, using the same pretrained FFBA weights without retraining or fine-tuning.

In-domain

GFMMethodPixel ↓\downarrowAngular ↓\downarrowATE ↓\downarrowRPE-t ↓\downarrowRPE-R ↓\downarrowAUC@1∘1^\circ ↑\uparrowAUC@3∘3^\circ ↑\uparrowAUC@5∘5^\circ ↑\uparrow
MapAnythingInit2.2833‾\underline{2.2833}0.1652‾\underline{0.1652}0.1249‾\underline{0.1249}0.1689‾\underline{0.1689}1.2900‾\underline{1.2900}6.12‾\underline{6.12}25.57‾\underline{25.57}39.23‾\underline{39.23}
MapAnythingFFBA0.89250.07170.10280.13740.973115.5642.1055.92
DA3-LargeInit3.0832‾\underline{3.0832}0.3020‾\underline{0.3020}0.0485‾\underline{0.0485}0.0581‾\underline{0.0581}0.5677‾\underline{0.5677}16.56‾\underline{16.56}40.22‾\underline{40.22}49.19‾\underline{49.19}
DA3-LargeFFBA2.22860.11630.04530.05680.417821.4043.4552.54

Cross-domain

GFMMethodPixel ↓\downarrowAngular ↓\downarrowATE ↓\downarrowRPE-t ↓\downarrowRPE-R ↓\downarrowAUC@1∘1^\circ ↑\uparrowAUC@3∘3^\circ ↑\uparrowAUC@5∘5^\circ ↑\uparrow
MapAnythingInit3.0423‾\underline{3.0423}0.2145‾\underline{0.2145}0.3005‾\underline{0.3005}0.1920‾\underline{0.1920}1.6559‾\underline{1.6559}4.09‾\underline{4.09}25.67‾\underline{25.67}40.74‾\underline{40.74}
MapAnythingFFBA1.45670.10490.26820.15211.50908.6935.6350.39
DA3-LargeInit2.5645‾\underline{2.5645}0.1789‾\underline{0.1789}0.2291‾\underline{0.2291}0.1814‾\underline{0.1814}0.9361‾\underline{0.9361}15.34‾\underline{15.34}48.23‾\underline{48.23}61.90‾\underline{61.90}
DA3-LargeFFBA1.06610.07570.18570.12940.813618.9352.5966.11

FFBA improves all reported aggregate metrics for both models, including when both the initializer and scene domain are unseen. These results demonstrate strong transfer across the tested GFMs and scene domains, showing that changing the upstream GFM does not necessarily require retraining. Full comparisons with BA and EPO are provided in App. H, Tabs. 11–12.


Unfinished. Please keep reading the comments.


4. Future role: standalone refinement, training-time supervision, or integration with GFMs

We appreciate your perspective on the future role of learned geometric refinement. We see learned geometric refinement as both a standalone module and a potential component of GFM training. Enforcing geometric consistency during training does not guarantee that predictions for every test scene agree with the observed image correspondences. A standalone refinement module therefore remains useful for correcting residual errors, particularly on new scenes or domains, as demonstrated by our unseen-GFM experiments.

The low refinement cost of FFBA also makes training-time geometric supervision a promising direction. Refined reconstructions could provide additional teacher targets for an upstream GFM, allowing it to learn corrections that would otherwise be applied after inference. Alternatively, refinement could be integrated into the GFM and share its feature representations. These approaches could reduce the need for a separate refinement stage at deployment.

An outstanding practical challenge is to acquire correspondences efficiently and reliably. Whether tracks come from an external matcher or a GFM, they should provide constraints supported by image evidence and avoid reinforcing errors in the predicted geometry. We therefore view FFBA as a step toward inexpensive geometric correction. Efficient correspondence construction is central to both its current standalone use and future integration into GFM training, and will be a major focus of our future work.


We appreciate your time and remain available to answer any further questions.

Reviewer_w8GAOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Feed-Forward Bundle Adjustment (FFBA), a method for refining the output of a Geometric Foundation Model (GFM), such as VGGT. To this end, the authors propose to train a neural network that takes as input an initial 3D reconstruction from a GFM (i.e. camera poses and 3D points), along with 2D tracks obtained through a separate, classical multi-view track construction procedure, and outputs a refined 3D reconstruction. The 2D tracks remain unchanged during the forward pass. Therefore, FFBA is directly comparable to standard Bundle Adjustment (BA), which also refines an initial 3D reconstruction using 2D tracks.

The proposed architecture is a geometry-conditioned graph neural network that essentially replaces the normal equations (used in classical BA) with message passing. FFBA is trained by minimizing a confidence-weighted reprojection error, thus it does not require 3D ground-truth.

FFBA is trained for 9.7 days on 8 A100 and evaluated on several benchmarks against BA and EPO, considering both in-domain and cross-domain datasets, with reprojection errors and camera pose accuracy as evaluation metrics. FFBA consistently improves upon the initial GFM reconstruction and achieves overall performance comparable to BA and EPO. In terms of efficiency, the refinement step is substantially faster than BA and EPO. However, the overall computational cost is dominated by the 2D track construction procedure, resulting in end-to-end inference times that are not substantially lower than those of BA and EPO in the reported experiments.

Critical strengths
  1. Non-trivial architecture and efficient refinement -- FFBA introduces a non-trivial neural network architecture that enables highly efficient inference, reducing the optimization time by more than an order of magnitude compared to classical BA.

  2. Clear presentation -- The method is well presented, and the paper is clearly written and easy to follow.

  3. Extensive experiments and balanced analysis -- The paper provides a substantial experimental evaluation across multiple benchmarks, including both in-domain and cross-domain settings. The analysis of the results is balanced and acknowledges cases where FFBA does not outperform the baselines.

Critical weaknesses
  1. Limited end-to-end computational benefits -- FFBA relies on 2D tracks whose construction dominates the overall computational cost. Consequently, although the refinement step itself is substantially faster than classical BA, the end-to-end inference time of FFBA is not significantly lower than that of BA or EPO in the reported experiments. I consider this a major limitation, particularly given the substantial training cost of FFBA (9.7 days on 8 GPUs). While the proposed architecture is interesting, the limited practical speedup raises questions about whether the benefits justify the training cost and additional complexity. This is my main concern and the primary reason why I am hesitant between a weak accept and a weak reject.

  2. Insufficient discussion of scalability and limitations -- I believe the previous point deserves a more thorough discussion in the limitations section. At the end of the conclusion, the authors suggest that the computational advantages of FFBA are expected to increase for large-scale scenes. However, this claim is not sufficiently justified in my opinion, as the cost of constructing 2D tracks is also expected to increase with scene size. It is therefore unclear whether the end-to-end computational benefits of FFBA would become more significant in such settings.

  3. Inconsistent improvements over classical BA -- Although FFBA consistently improves upon the initial GFM reconstruction, it does not consistently outperform classical BA, even on in-domain datasets. For instance, on ScanNet++ (Table 1), BA achieves better results. This raises questions about the extent to which FFBA can effectively replace classical BA, particularly in settings where the latter already performs well.

Questions for authors
  1. Graph attention network details (Line 269) -- The paper states that FFBA uses graph attention networks, but the exact formulation is not provided. Could the authors include the corresponding equations, perhaps in an appendix? I expect these equations to be relatively straightforward, and providing them would make this part of the architecture easier to understand.

  2. Layer-specific focal decoder (Line 287) -- The authors mention that the focal decoder is layer-specific. What motivates this design choice? Would sharing the decoder across layers affect performance?

  3. End-to-end efficiency and scalability -- Could the authors elaborate on the concerns raised in Weaknesses #1 and #2? In particular, I would be interested in understanding under which conditions FFBA is expected to provide a substantial end-to-end computational advantage over classical BA, rather than merely accelerating the refinement step. It would also be helpful to provide additional evidence or analysis supporting the claim that these advantages should become more pronounced for larger scenes.

4 comments

We appreciate the reviewer's thorough assessment and helpful feedback. Below, we report additional efficiency and large-scene experiments, discuss accuracy trade-offs, and clarify the network design to provide a clearer understanding of the value and limitations of FFBA.


1. End-to-end efficiency, frontend bottlenecks, and training cost (Weakness 1 / Question 3)

We share this concern: the substantial speedup in the refinement stage translates into more modest end-to-end gains due to frontend data-preparation costs. Track construction is the main frontend bottleneck: matching and geometric verification account for 78.2% of track-construction time on cross-domain data. To reduce frontend overhead, we first accelerated 3D point initialization through batched triangulation operations. This implementation change applies to both FFBA and BA, preserving candidate pairs, filtering, and optimization settings. Updated results are below (Tabs. 3 and 8).

ScopeMethodData preparation (s) ↓\downarrowOptimization (s) ↓\downarrow# Iter.Total (s) ↓\downarrow
In-domainBA6.352.59‾\underline{2.59}698.94‾\underline{8.94}
In-domainEPO3.6518.82136022.47
In-domainFFBA6.28‾\underline{6.28}0.07—6.36
Cross-domainBA7.832.10‾\underline{2.10}529.93‾\underline{9.93}
Cross-domainEPO4.0311.0896615.11
Cross-domainFFBA7.71‾\underline{7.71}0.09—7.81

FFBA reduces total runtime relative to BA by approximately 29% in-domain and 21% cross-domain, and is faster overall than EPO in both settings.

We also evaluated sparser candidate view graphs to reduce the cost of matching and geometric verification. Our conservative default takes the union of ALIKED top-10 retrieval neighbors and pairs whose sequence indices differ by at most five, then removes duplicates. The following alternatives reduce the matching workload (App. F, Tab. 9).

Pair selection# PairsTracks (s) ↓\downarrowTotal (s) ↓\downarrowAUC@1∘1^\circ (%) ↑\uparrowAUC@3∘3^\circ (%) ↑\uparrowAUC@5∘5^\circ (%) ↑\uparrow
ALIKED-10 + W5(Default)297.35.047.8122.7557.2069.64
ALIKED-10241.14.306.6722.6257.0769.52
Window-5178.33.37‾\underline{3.37}6.07‾\underline{6.07}22.0456.6369.28
SALAD-10235.14.577.2122.74‾\underline{22.74}57.18‾\underline{57.18}69.62‾\underline{69.62}
SALAD-5118.62.765.0622.3756.9269.46

SALAD-5 reduces track-construction time by approximately 45%, with only a 0.18 percentage-point decrease in AUC@5. Thus, frontend costs can be reduced without changing FFBA.

Training takes 9.7 days on eight A100 GPUs, but this offline cost is shared across subsequent reconstructions. The same model is reused across scenes and unseen GFMs without fine-tuning (App. H). Its practical value therefore increases with repeated use. We will further scale up training and diversify geometric error patterns to improve the currently uneven gains on unseen distributions. We will also explore more efficient frontends to better translate fast FFBA refinement into end-to-end efficiency gains.


Unfinished. Please keep reading the comments.


2. Scalability to large scenes (Weakness 2 / Question 3)

We appreciate your suggestion to evaluate larger scenes. We evaluated five complete T&T scenes with 251–410 views, versus at most 96 during training. Results on A100 GPUs are averaged equally over these scenes and three initializers. Church and Courthouse are excluded for all methods because some initializers exceed GPU memory capacity.

Method / Pair selectionData prep. (s) ↓\downarrowOptim. (s) ↓\downarrowTotal (s) ↓\downarrowAUC@1∘1^\circ (%) ↑\uparrowAUC@3∘3^\circ (%) ↑\uparrowAUC@5∘5^\circ (%) ↑\uparrow
BA / Default hybrid137.2050.32187.5222.6268.8181.06
BA / SALAD-5110.1638.35148.5116.1361.3876.03
BA / SALAD-10129.1147.68176.7921.2067.2980.12
FFBA / Default hybrid135.190.70‾\underline{0.70}135.8931.2373.0383.51
FFBA / SALAD-5109.31‾\underline{109.31}0.66109.9629.7872.1382.92
FFBA / SALAD-10126.730.71127.4430.92‾\underline{30.92}72.93‾\underline{72.93}83.44‾\underline{83.44}
EPO / Native pipeline81.5232.42113.94‾\underline{113.94}24.7469.1281.16

With the default strategy, FFBA achieves approximately 72×72\times faster optimization and 27.5% lower total runtime than BA, with higher AUC at all three thresholds. Default FFBA remains slower overall than EPO; SALAD-5 makes it slightly faster while retaining higher AUC, with only a 0.59 percentage-point AUC@5 decrease from default FFBA.

We agree that the original scaling claim was too broad: FFBA retains effective refinement and a substantial optimization-time advantage on these large scenes, while end-to-end gains depend on frontend costs, not scene size alone (App. G, Tab. 10).


3. Inconsistent accuracy advantages over BA (Weakness 3)

We agree that the accuracy advantage varies across distributions. On ScanNet++, FFBA improves AUC@5 from 17.75% to 40.56%, below BA's 54.69%; on DL3DV10K and MVImgNet, it outperforms BA and EPO on all reported camera metrics. ScanNet++ has harder initializations, lower complete-registration coverage with COLMAP, and fewer constructed samples (Tabs. 1, 5, and 13). Weak texture, limited parallax, and reduced long-range constraints after scene splitting may contribute, but their individual effects have not been isolated. Broader coverage of difficult examples and improved training distributions are therefore priorities for future work.


4. Graph-attention equations (Question 1)

We added “Graph Attention Network” in App. B, detailing GATv2 scoring, neighbor-wise normalization, weighted aggregation, and multi-head attention. We also added a reference to this section in the main text.


5. Layer-specific focal decoder (Question 2)

Focal length controls view-wide projection scale and can compensate for depth or pose errors. Pose and point heads predict increments to the current state, modulated by layer-specific learned step sizes. The focal branch instead re-estimates an overall correction relative to the initial focal length, without a separate step-size controller. Independent focal decoders therefore let each refinement stage learn how to extract calibration information from its current geometric features (Sec. 3.4, Eqs. 14–15).

We are also conducting an ablation study of the shared focal decoder; because training is time-consuming, we will include the results in the revised manuscript before the rebuttal deadline.


We appreciate your time and remain available to answer any further questions.

Dear authors,

Thank you for addressing my questions. I appreciate the care and candor with which you responded to the feedback. I have no further questions at this stage. I will finalize my assessment after discussing the paper with the other reviewers.

Best regards,

Thank you again for your thorough and insightful review. Your suggestions have helped us further improve our method, particularly its efficiency. We sincerely appreciate your time and careful consideration.

Reviewer_dN3qOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper describes a method to replace the BA with graph attention to work with VGGT/VGGT-Omega like fundation geometry model. On top of this, the author added uncertainty-guided filtering, and confidence-weighted loss.

The paper validated the idea of the combination of 3 datasets and reported results in pairs with the Colmap baseline.

I have printed this paper and read it three times before posting my comments.

Critical strengths

This paper wraps itself as a new way to work with fundation models (VGGT/VGGT-Omega) and to add learnable BA layer on the SfM pipeline. The motivation itself is very strong and its application is pertinent enough to draw a large audience.

The authors created a near 1M dataset to train this network, which has significance for the fast-changing 3D community. I see values from this line of research. The authors also compare the method a good baseline, such as Colmap BA.

Critical weaknesses

I will say that this paper is misleading in the way they present and is overstated from my perspective. The reading is particularly difficult, specifically for readers without hands-on SfM experiences.

The core setting of this paper is that they used a VGGT-like model to retrieve the intrinsic+camera poses. They then use these two cues, coupled with 2D correspondings (from optical flow or from feature matching? I didn't find this info from the paper), to create sparse 3D points, similar to triangulation. This becomes their problems to be solved.

I have personally worked on these fundation models for a while and it is very clear that the prediction from these models are particularly noisy. Intrinsic values (focal length can have a std of up to 20%) and the authors assume images are indepedent. I must understand that the "input" of the problem is particularly bad. There is no discussion, no visualization, and no data on this important fact.

Naturally, you will also need a baseline of running colmap from end to end, instead of just running BA to fairly evaluate your setting and find from which side your indeed add values.

Section 3.3 is particularly hard to follow and evaluate. I understand they cite one paper (They must cite it at line 179 p4) and do extension. The way they present it should be more self-contained.

I am also particularly curious on one more setting, given colmap generated (near ground truth) inputs, how the proposed algorithm will react.

Questions for authors

I have previously posed a few questions.

  1. a natural comparison with Colmap GT, so that we understand to which part, to what extend you are trying to solve this problem.
  2. what would be the reaction if your input is perfect.
  3. could you please improve the readability of this paper by including more explanations and figures?
2 comments

We thank the reviewer for the careful reading and constructive suggestions. Below, we address your concerns about initialization quality through the added COLMAP comparisons and clarify our pipeline with expanded descriptions and new illustrations.


1. Comparison with complete COLMAP reconstruction

Thank you for suggesting a direct comparison with complete COLMAP reconstruction. We evaluated the complete COLMAP pipeline, including internal BA. It fully registers 2,898/3,132 in-domain and 172/193 cross-domain image sets. The tables below compare its results with GFM initializations and GFM + FFBA; COLMAP + FFBA addresses the next question.

In-domain results.

MethodPixel ↓\downarrowAngular ↓\downarrowATE ↓\downarrowRPE-t ↓\downarrowRPE-R ↓\downarrowAUC@1∘1^\circ ↑\uparrowAUC@3∘3^\circ ↑\uparrowAUC@5∘5^\circ ↑\uparrow
GFM (Init)1.62430.15260.0464‾\underline{0.0464}0.0616‾\underline{0.0616}0.372713.9939.0951.26
GFM + FFBA0.65570.06080.03980.05550.288721.3449.5162.10
COLMAP0.5094‾\underline{0.5094}0.0476‾\underline{0.0476}0.05730.06820.314127.80‾\underline{27.80}52.77‾\underline{52.77}62.93‾\underline{62.93}
COLMAP + FFBA0.49390.04580.05670.06710.3112‾\underline{0.3112}28.1153.5263.67

Cross-domain results.

MethodPixel ↓\downarrowAngular ↓\downarrowATE ↓\downarrowRPE-t ↓\downarrowRPE-R ↓\downarrowAUC@1∘1^\circ ↑\uparrowAUC@3∘3^\circ ↑\uparrowAUC@5∘5^\circ ↑\uparrow
GFM (Init)1.93620.13750.15140.1159‾\underline{0.1159}0.6391‾\underline{0.6391}16.0249.7163.94
GFM + FFBA0.75910.05400.13430.09460.568422.7557.2069.64
COLMAP0.4925‾\underline{0.4925}0.0339‾\underline{0.0339}0.12640.15280.810446.9373.9381.48
COLMAP + FFBA0.47850.03300.1275‾\underline{0.1275}0.15350.809345.87‾\underline{45.87}73.28‾\underline{73.28}81.02‾\underline{81.02}

In the reported averages, GFM + FFBA approaches COLMAP's in-domain AUC@5 (62.10% vs. 62.93%) while achieving lower ATE and RPE. Cross-domain, COLMAP retains higher AUC, whereas GFM + FFBA has lower RPE. These results show that FFBA improves GFM initializations and narrows their AUC gap to complete COLMAP reconstruction, while remaining competitive on several pose-error metrics. COLMAP also relies on reliable image correspondences, and some image sets in our evaluation are not fully registered. Evaluating complete reconstruction pipelines therefore requires considering both geometric accuracy and registration coverage. Detailed results and registration counts for each dataset are provided in Tabs. 1–2 of the main paper and App. I, Tabs. 13–14.


2. Refining COLMAP initial reconstructions

We appreciate your suggestion to test FFBA on reconstructions already optimized by COLMAP. We applied the same pretrained FFBA to COLMAP outputs, comparing both methods on identical fully registered samples. The tables in our response to the previous question summarize the average results of COLMAP and COLMAP + FFBA on the in-domain and cross-domain test sets. In-domain, COLMAP + FFBA improves all eight average metrics. Cross-domain, most average metrics deteriorate slightly, while some still improve over COLMAP.

These results show that FFBA can also refine COLMAP initializations, although its benefits vary with the data distribution and input geometric quality. COLMAP outputs have already undergone scene-level optimization, leaving less room for further improvement, and their residual error patterns may differ from those of the GFM initializations used during training. These factors may partly explain the slight degradation in some cross-domain metrics after FFBA refinement. We believe that training on data covering more initialization sources and error magnitudes could better balance error correction and the preservation of high-precision geometry. Further discussion is provided in App. I.


Unfinished. Please keep reading the comments.


3. Clarifying the source of correspondences and improving readability

Thank you for highlighting the need for a clearer description. Our correspondences come from feature matching, not optical flow: ALIKED features are matched by LightGlue, verified with MAGSAC++, and merged into tracks by conflict-aware union–find. FFBA keeps measured 2D coordinates fixed while refining cameras and 3D points and masking unreliable observations.

We have expanded Sec. 4 and added Figs. 3–4 to explain sample compilation and track construction. In Sec. 3.3, we explain the graph's node types and connections and explicitly cite GASfM when introducing the heterogeneous graph organization. Apps. A–B further provide refinement pseudocode and the attention equations, making the method description more self-contained and easier to follow.


4. GFM initialization quality

Our main results tables report the GFM initializations using camera-pose and reprojection-error metrics. We acknowledge that these initializations still lag behind complete COLMAP reconstruction on metrics such as relative-pose AUC. Our goal is precisely to narrow this gap through global geometric refinement that jointly corrects camera parameters and 3D structure, serving a role analogous to BA in a traditional SfM pipeline. Following your suggestion, we added a complete COLMAP baseline to assess the reconstruction quality of GFM initializations before and after FFBA refinement. We also evaluated COLMAP + FFBA to examine its behavior on high-quality inputs that have already undergone classical optimization. Both experiments and their findings are discussed in our preceding responses.


We appreciate your time and remain available to answer any further questions.

Comments on the paper

1 comment

We thank the reviewers for their careful evaluation of our work and their constructive feedback. Reviewers highlighted the strong motivation and the value of the shared-track dataset (dN3q), as well as the non-trivial network design, highly efficient refinement, extensive evaluation, and balanced analysis (w8GA). They also recognized the feasibility of replacing per-scene iterative BA with learned feed-forward refinement and the potential of this direction for future research (CQ21). Following their suggestions, we have added experiments to address each reviewer's main concern.

  • Comparison with complete COLMAP reconstruction (Reviewer dN3q). We added an evaluation of the complete COLMAP pipeline, covering reconstruction accuracy and image registration coverage. FFBA improves the initial GFM reconstructions to varying degrees, bringing the average in-domain AUC@5 close to that of COLMAP. We also applied the same FFBA model to COLMAP reconstructions that had already undergone BA. We observed small improvements in all aggregate in-domain metrics and slight declines in most aggregate cross-domain metrics. Our analysis suggests that these declines may be related to substantial differences in the quality and distribution of the initial reconstructions.

  • End-to-end efficiency and large-scene performance (Reviewer w8GA). We applied batched triangulation to both FFBA and BA to improve the efficiency of the pipeline implementation. In the updated in-domain and cross-domain evaluations, FFBA remains more than an order of magnitude faster than BA during refinement, with approximately 29% and 21% lower end-to-end runtime, respectively. On five complete T&T scenes, each containing 251–410 views, FFBA also maintains its refinement speed advantage, while end-to-end gains still depend on the cost of frontend matching and geometric verification. Our future work will therefore focus primarily on optimizing the FFBA frontend.

  • Generalization to unseen initializers (Reviewer CQ21). We directly applied the same FFBA model to MapAnything and DA3-Large, neither of which was used during training, and evaluated it on three in-domain and four cross-domain datasets. Without retraining or fine-tuning, FFBA improves all reported metrics for both initializers, including settings in which both the GFM initializer and the scene domain were unseen during training. These results show that the geometric update rules learned by FFBA can transfer to new upstream GFMs. Changing the initializer therefore does not necessarily require retraining, supporting the use of FFBA as a reusable geometric refinement module.

These experiments have been incorporated into the revised manuscript. We have also addressed the source of 2D correspondences and the clarity of the method description (Reviewer dN3q), track-construction overhead, training cost, and graph-attention equations (Reviewer w8GA), the accuracy gap relative to classical BA (Reviewers w8GA and CQ21), and dependence on training data and integration with upstream GFMs (Reviewer CQ21). Regarding the layer-specific focal decoder design (Reviewer w8GA), we have completed the ablation with a shared focal decoder and updated Table 4. The revision is now complete, and all additions and changes are marked in red for ease of reference. We thank the reviewers again for their careful assessment and valuable suggestions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.