- Rating
- 3
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper introduces GeoMVT, a depth-free multi-view point tracking framework for accurate 2D tracking and 3D motion recovery. It proposes an epipolar-guided geometric transform attention mechanism for cross-view correlation aggregation and a geometry-guided transformer for iterative trajectory refinement, enabling 3D trajectory recovery through weighted triangulation. Experiments on four benchmarks show that GeoMVT consistently outperforms existing multi-view 2D trackers and achieves better 3D tracking results than MVTracker on DexYCB and Harmony4D, without requiring depth input.
Critical strengths
-
Practical Relevance. The proposed method enables multi-view 2D point tracking and 3D trajectory recovery without requiring depth input, which is useful for applications such as robotics and 3D motion reconstruction.
-
Strong Empirical Performance. GeoMVT consistently outperforms existing multi-view 2D trackers on all four benchmarks. It also achieves better 3D tracking results than MVTracker on DexYCB and Harmony4D without using depth input.
-
Comprehensive Ablation Studies. The paper provides extensive ablation studies to examine the effectiveness of the proposed components.
Critical weaknesses
-
Unclear Contribution Attribution. EpiGTA builds on existing geometric attention and epipolar constraints. In Table 4, adding GTA to epipolar-guided attention yields only marginal gains. Although the full framework performs well, the individual contributions of geometric transformation, epipolar guidance, and the two-stage design remain insufficiently distinguished.
-
Insufficient Direct Evidence for Improved 3D Geometric Consistency. The paper claims that geometry-guided reasoning improves cross-view consistency and leads to more accurate 3D tracking. While Figure 7 provide supporting evidence by comparing 3D errors within matched 2D error bins, direct geometric consistency metrics, such as epipolar or reprojection errors, are still missing. Moreover, the ablation studies primarily focus on 2D tracking, leaving the individual contributions of CCA and GIU to 3D performance unclear. Additional 3D ablations and evaluations of cross-view geometric consistency would help substantiate the paper's central claims.
-
Reliance on Cross-View Correspondences and Accurate Camera Calibration. The method assumes known point correspondences across cameras and accurate camera calibration, which may be difficult to obtain in real-world scenarios. The paper does not evaluate robustness to incorrect cross-view correspondences.
Questions for authors
-
Could the authors further clarify the individual contribution of GTA-based geometric transformations? Table 4 shows only marginal gains from adding GTA to epipolar-guided attention. It would be helpful to evaluate a full-model variant that retains epipolar guidance in CCA but replaces GTA-based transformations with standard cross-view attention, ablating the geometric transformations in CCA and GIU separately. This would help better isolate the contributions of the proposed geometric reasoning mechanisms.
-
Could the authors provide more direct evidence that GeoMVT improves cross-view geometric consistency? While Figure 7 and Appendix B.9 show lower 3D errors within matched 2D error bins, additional metrics such as epipolar errors or reprojection errors would help substantiate this claim. Moreover, reporting 3D AJ and MTE for the CCA-only, GIU-only, and full-model variants would clarify how each component contributes to 3D tracking performance.
-
The method assumes that query points corresponding to the same physical point are already identified across camera views. How would GeoMVT behave when these cross-view correspondences are imperfect or partially incorrect? An evaluation with perturbed or mismatched cross-view correspondences would help clarify the robustness of the proposed framework and its applicability to practical scenarios.
-
Appendix B.1 provides a useful analysis of robustness to camera calibration errors in terms of 2D tracking accuracy. However, calibration errors may also directly affect the subsequent 3D triangulation. Could the authors further evaluate 3D tracking performance under calibration perturbations, particularly in terms of 3D AJ and MTE? This would help clarify how sensitive the reported 3D improvements are to calibration accuracy.
3 comments
Response to Reviewer xQKa (Part 2/3)
W3 & Q3: Robustness to incorrect cross-view correspondences
Thank you for this insightful suggestion. To evaluate realistic initialization errors, we retain the ground-truth query in the first valid view and perturb queries in the remaining valid views by 1–3 px. Table R3 (Appendix Table B.21) shows that both methods lose accuracy as correspondence errors increase, but GeoMVT consistently outperforms MV-TAP [1] across all three metrics on all four datasets. Even at 2 px, GeoMVT retains 93.1–95.0% of its unperturbed AJ. These results support robustness to imperfect initial cross-view correspondences without requiring accurate queries in every view. This robustness suggests that GeoMVT could be paired with recent correspondence methods such as RoMa v2 [2] or VGGT [3] to initialize queries across views and enable accurate multi-view tracking.
Table R3. 2D tracking with imperfect initial cross-view correspondences.
$ \scriptsize\begin{array}{c|c|c|c|c|c}\text{Error (px)}&\text{Method}&\text{Panoptic Studio}&\text{DexYCB}&\text{MV-Kubric}&\text{Harmony4D}\\&&\begin{array}{ccc}\text{AJ}&\delta_{\mathrm{avg}}&\text{OA}\end{array}&\begin{array}{ccc}\text{AJ}&\delta_{\mathrm{avg}}&\text{OA}\end{array}&\begin{array}{ccc}\text{AJ}&\delta_{\mathrm{avg}}&\text{OA}\end{array}&\begin{array}{ccc}\text{AJ}&\delta_{\mathrm{avg}}&\text{OA}\end{array}\\\hline 0&\begin{array}{l}\text{MV-TAP}\\\text{Ours}\end{array}&\begin{array}{ccc}43.5&67.2&74.8\\\mathbf{50.4}&\mathbf{72.2}&\mathbf{81.4}\end{array}&\begin{array}{ccc}44.9&62.7&77.7\\\mathbf{49.8}&\mathbf{68.5}&\mathbf{83.4}\end{array}&\begin{array}{ccc}49.5&68.8&81.9\\\mathbf{58.0}&\mathbf{77.2}&\mathbf{86.9}\end{array}&\begin{array}{ccc}38.9&71.3&69.3\\\mathbf{44.4}&\mathbf{75.0}&\mathbf{72.3}\end{array}\\\hline 1&\begin{array}{l}\text{MV-TAP}\\\text{Ours}\end{array}&\begin{array}{ccc}42.6&66.4&74.5\\\mathbf{49.8}&\mathbf{71.5}&\mathbf{81.3}\end{array}&\begin{array}{ccc}44.4&62.1&77.8\\\mathbf{49.2}&\mathbf{67.9}&\mathbf{83.3}\end{array}&\begin{array}{ccc}48.6&67.9&81.9\\\mathbf{57.1}&\mathbf{76.3}&\mathbf{86.8}\end{array}&\begin{array}{ccc}38.1&70.3&69.2\\\mathbf{43.7}&\mathbf{74.2}&\mathbf{72.1}\end{array}\\\hline 2&\begin{array}{l}\text{MV-TAP}\\\text{Ours}\end{array}&\begin{array}{ccc}40.5&63.8&74.2\\\mathbf{47.9}&\mathbf{69.0}&\mathbf{81.1}\end{array}&\begin{array}{ccc}41.5&58.5&77.5\\\mathbf{46.5}&\mathbf{64.2}&\mathbf{83.2}\end{array}&\begin{array}{ccc}45.3&64.0&81.3\\\mathbf{54.0}&\mathbf{72.8}&\mathbf{86.5}\end{array}&\begin{array}{ccc}35.9&67.5&69.0\\\mathbf{41.8}&\mathbf{71.5}&\mathbf{71.9}\end{array}\\\hline 3&\begin{array}{l}\text{MV-TAP}\\\text{Ours}\end{array}&\begin{array}{ccc}37.8&60.5&73.6\\\mathbf{45.3}&\mathbf{66.0}&\mathbf{80.6}\end{array}&\begin{array}{ccc}39.1&55.5&77.3\\\mathbf{43.9}&\mathbf{61.1}&\mathbf{82.8}\end{array}&\begin{array}{ccc}41.5&59.9&80.5\\\mathbf{50.4}&\mathbf{68.8}&\mathbf{85.9}\end{array}&\begin{array}{ccc}33.2&63.9&68.7\\\mathbf{39.0}&\mathbf{68.2}&\mathbf{71.4}\end{array}\\\hline \end{array} $
References.
- [1] MV-TAP: Tracking Any Point in Multi-View Videos, CVPR 2026.
- [2] RoMa v2: Harder Better Faster Denser Feature Matching, ECCV 2026.
- [3] VGGT: Visual Geometry Grounded Transformer, CVPR 2025.
Response to Reviewer xQKa (Part 3/3)
Q4: 3D tracking performance under calibration perturbations
Thank you for this insightful suggestion. We separately perturb camera rotations with fixed centers and camera positions with fixed rotations, applying the perturbed extrinsics to both tracking and triangulation.
(i) Rotation perturbations. Table R4 shows that both methods lose AJ and incur larger MTE as rotation errors increase. Nevertheless, GeoMVT retains higher AJ and lower MTE than MV-TAP [1] across all four datasets throughout the tested range.
(ii) Translation perturbations. Table R5 shows that our advantage also persists throughout the tested translation-error range. These results demonstrate that our geometric design retains its comparative benefit under calibration perturbations, although inaccurate cameras affect both tracking and 3D recovery.
The corresponding curves are provided in Appendix Figs. B.4–B.7.
Table R4. 3D tracking under camera rotation perturbations.
Table R5. 3D tracking under camera translation perturbations.
\scriptsize\begin{array}{c|c|c|c|c}\text{Translation error (\\%)}&\text{Panoptic Studio}&\text{DexYCB}&\text{MV-Kubric}&\text{Harmony4D}\\\\\hline \begin{array}{c}\phantom{0}\\\\\phantom{0}\\\\0\\\\0.25\\\\0.5\\\\1\\\\2\end{array}&\begin{array}{c|c}\text{MV-TAP}&\text{Ours}\\\\\begin{array}{cc}\text{AJ}&\text{MTE}\\\\\hline 81.6&5.1\\\\81.3&5.2\\\\80.8&5.4\\\\79.2&6.0\\\\72.1&7.4\end{array}&\begin{array}{cc}\text{AJ}&\text{MTE}\\\\\hline \mathbf{90.1}&\mathbf{3.1}\\\\\mathbf{89.9}&\mathbf{3.3}\\\\\mathbf{89.5}&\mathbf{3.5}\\\\\mathbf{87.8}&\mathbf{4.2}\\\\\mathbf{78.6}&\mathbf{6.0}\end{array}\end{array}&\begin{array}{c|c}\text{MV-TAP}&\text{Ours}\\\\\begin{array}{cc}\text{AJ}&\text{MTE}\\\\\hline 69.4&2.0\\\\69.4&2.0\\\\69.5&2.0\\\\69.2&2.1\\\\65.9&2.3\end{array}&\begin{array}{cc}\text{AJ}&\text{MTE}\\\\\hline \mathbf{78.4}&\mathbf{1.2}\\\\\mathbf{78.2}&\mathbf{1.3}\\\\\mathbf{77.7}&\mathbf{1.3}\\\\\mathbf{76.3}&\mathbf{1.5}\\\\\mathbf{70.4}&\mathbf{1.8}\end{array}\end{array}&\begin{array}{c|c}\text{MV-TAP}&\text{Ours}\\\\\begin{array}{cc}\text{AJ}&\text{MTE}\\\\\hline 29.4&10.7\\\\29.3&10.7\\\\28.6&10.7\\\\26.7&11.0\\\\22.8&11.7\end{array}&\begin{array}{cc}\text{AJ}&\text{MTE}\\\\\hline \mathbf{48.4}&\mathbf{3.7}\\\\\mathbf{47.9}&\mathbf{3.8}\\\\\mathbf{46.1}&\mathbf{3.9}\\\\\mathbf{41.4}&\mathbf{4.3}\\\\\mathbf{33.1}&\mathbf{5.4}\end{array}\end{array}&\begin{array}{c|c}\text{MV-TAP}&\text{Ours}\\\\\begin{array}{cc}\text{AJ}&\text{MTE}\\\\\hline 55.1&12.8\\\\55.1&12.8\\\\54.7&12.9\\\\53.5&13.2\\\\49.3&14.6\end{array}&\begin{array}{cc}\text{AJ}&\text{MTE}\\\\\hline \mathbf{63.6}&\mathbf{9.2}\\\\\mathbf{63.3}&\mathbf{9.2}\\\\\mathbf{62.9}&\mathbf{9.3}\\\\\mathbf{60.7}&\mathbf{10.2}\\\\\mathbf{54.7}&\mathbf{11.9}\end{array}\end{array}\end{array}
References.
- [1] MV-TAP: Tracking Any Point in Multi-View Videos, CVPR 2026.
We sincerely thank you again for the valuable comments provided during this review.
Response to Reviewer xQKa (Part 1/3)
W1 & Q1: Detailed contribution ablations
Thank you for this suggestion. We have expanded the component ablations with a variant that retains epipolar guidance in CCA but removes its geometric transformations, while keeping GIU unchanged. Table R1 (Table 3 in the manuscript) reports both 2D and 3D tracking results. Adding geometric transformations in CCA further improves the epipolar-guided variant, while coupling CCA with GIU yields higher AJ than either stage alone. Together, the 2D and 3D results further support the effectiveness of geometric transformations, epipolar guidance, and the coupled two-stage design.
Table R1. Ablation of the main components on 2D and 3D point tracking. Additional results is in appendix Table B.3.
CCAGTA0×✓××✓EPI0×✓×✓✓GIU00××✓✓✓2DPanoptic StudioAJ44.046.248.749.650.9δavg67.169.270.571.472.4OA75.276.880.081.081.5DexYCBAJ46.449.049.751.351.4δavg63.666.967.569.069.0OA78.581.483.684.984.43DPanoptic StudioAJ82.685.688.389.590.1δavg87.790.692.593.493.8OA97.196.397.697.797.9MTE4.94.13.63.33.1DexYCBAJ70.774.476.977.678.4δavg79.082.884.785.686.3OA92.992.693.093.093.0MTE1.81.51.41.31.2
W2.2 & Q2.2: Additional 3D ablations
We have extended Table R1 with 3D tracking metrics. The 3D results largely mirror the 2D trends, with the coupled model achieving the highest AJ and lowest MTE on both datasets. Together with the reprojection analysis, these results further support the effectiveness of our geometric design in improving cross-view consistency and 3D trajectory recovery.
W2.1 & Q2.1: Additional evidence of cross-view geometric consistency
Thank you for this suggestion. We triangulate 2D tracks using the same weighted procedure and measure mean pixel distances between their reprojections and predicted 2D positions. Table R2 (Appendix Table B.20) shows that GeoMVT consistently achieves lower reprojection error than MV-TAP [1] in every available bin across all four datasets. Our multi-view tracks therefore agree more closely with a common 3D trajectory, providing further evidence of improved cross-view geometric consistency.
Table R2. Reprojection errors within matched 2D error bins. 2D error is the mean pixel error at ground-truth visible positions.
2D error0[0,1)[1,2)[2,3)[3,4)[4,5)[5,6)[6,7)[7,8)Panoptic StudioMV-TAP1.182.182.953.314.423.925.826.06Ours0.751.461.982.442.932.753.763.64Δ−0.43−0.72−0.97−0.87−1.49−1.17−2.06−2.42DexYCBMV-TAP0.611.872.524.025.105.086.816.21Ours0.411.211.752.673.283.734.874.45Δ−0.20−0.66−0.77−1.34−1.82−1.36−1.94−1.76MV-KubricMV-TAP3.975.889.4025.586.7712.477.789.22Ours1.622.272.663.353.295.364.535.85Δ−2.34−3.61−6.74−22.23−3.48−7.12−3.24−3.37Harmony4DMV-TAP−1.141.982.583.463.694.655.99Ours−1.041.561.922.422.762.963.23Δ−−0.11−0.43−0.66−1.04−0.93−1.69−2.75
References.