acceptodds
Under review as a conference paper at ICLR 2027

ConTraPosE: Geometry-Aware Contrast Maximization for Event-Based Object Tracking

Abstract

Accurate 6DoF object tracking under rapid motion remains challenging for frame-based cameras due to motion blur and limited temporal resolution. Event cameras alleviate these limitations, but their sparse asynchronous measurements provide little direct appearance information, making conventional appearance-based pose estimation difficult. We introduce ConTraPosE, a tracking method based on a geometry-aware contrast formulation for estimating 6DoF object motion directly from raw events. Given a known CAD model and a candidate pose trajectory, we use the object geometry to construct its induced image-plane motion field, warp the observed events accordingly, and evaluate the resulting alignment with a normalized contrast objective. This provides a direct pose-update signal without image reconstruction or explicit event-to-object correspondences. We integrate this objective into continuous event-only tracking with adaptive temporal windows and optionally incorporate depth as an additional geometric constraint for absolute pose correction. Experiments on Event6D and SPADES show strong performance in both event-only and event–depth settings, with consistent robustness across different object geometries, appearances, motion regimes, event corruption, occlusion, and initialization perturbations. These results demonstrate that geometry-induced motion consistency provides an effective alternative to appearance-based comparison for event-driven 6DoF object tracking.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Event camera, pose estimation, pose tracking, 6 DoF

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2m6LOpenReview ↗
Rating
1(−1)
Confidence
2
Informativeness
1(−1)
Presentation
1
Soundness
2
Read review
Paper summary

ConTraPosE extends contrast maximization to event-based 6DoF object tracking using a known CAD model and initial pose. It evaluates candidate motions through geometry-induced event warping and normalized contrast, with adaptive temporal windows and optional depth correction. Experiments on Event6D and SPADES evaluate tracking accuracy, robustness, and computational cost.

Critical strengths
  1. The method provides a clear motion-consistency formulation for event-based object tracking, reducing dependence on appearance matching.
  2. Experiments on two datasets demonstrate strong tracking performance, complemented by robustness evaluations under varied conditions.
Critical weaknesses
  1. Motion compensation and IWE contrast maximization without additional appearance information or explicit event correspondences are central to [1]. Meanwhile, [2] already use known 3D models and rigid-motion projection with motion interpolation for event-based 6DoF object tracking, although their objective is brightness-increment alignment. ConTraPosE appears to combine geometry-based motion construction with the established CMax framework, but does not sufficiently establish the methodological advance of its specific design beyond this combination. This leaves the methodological contribution closer to a task-specific extension than a substantial new formulation.

[1] Gallego G, Rebecq H, Scaramuzza D. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation[C]//2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. IEEE, 2018: 3867-3876. [2] Kang Y, Caron G, Ishikawa R, et al. Direct 3D model-based object tracking with event camera by motion interpolation[C]//2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024: 2645-2651.

  1. The experiments demonstrate the effectiveness of the complete tracker, but do not isolate the contribution of the proposed core formulation. Comparisons with EDOPT and ETAP+PnP involve different tracking mechanisms, while the reported ablation primarily evaluates adaptive temporal windows. Controlled comparisons with CAD-based CMax variants under matched optimization and window settings are needed to establish the specific benefit of the proposed motion-field construction and normalized objective.

  2. The layout is overly compressed, with insufficient spacing around figures and tables. Some visual elements in Figure 2 overlap, while the crowded arrangement of figures, tables, and surrounding text reduces readability.

Questions for authors
  1. What specific methodological advance does the proposed formulation provide beyond combining the CMax framework in [1] with model-based rigid-motion projection and interpolation as used in [2]?
  2. Are there existing CMax-based methods that could provide a closer comparison in this tracking setting? If not, could the authors provide targeted ablations of motion-field fusion, support weighting, and contrast normalization, while keeping the motion parameterization, optimizer, and temporal-window policy fixed?
6 comments
AuthorsOpenReview ↗

Methodological Novelty: Motion Fields, Event Support, and Adaptive Windows (W1/Q1)

Positioning and methodological novelty (W1/Q1)

We thank the reviewer and agree that we should distinguish our methodological contributions more clearly from existing methods. CMax [1], CAD-based projection, and motion interpolation [2] are established, and we do not claim novelty in these components individually. Our contribution lies in adapting the CMax objective to CAD-based 6DoF object tracking, where both the image motion field and the spatial support of object events depend on the candidate pose trajectory. Simply combining conventional CMax with CAD-induced object motion, even with motion interpolation, does not fully address these challenges. Specifically, we identify and address two methodological challenges:

1. Accurate motion-field representation. Existing CMax formulations commonly approximate optical flow as a constant over short temporal windows. For object tracking, longer windows may be needed to accumulate sufficient observable motion, particularly for small or distant objects. However, the projected motion field and visible surfaces can change substantially within these windows. While BIAM [2] uses motion interpolation, its motion representation relies on fixed reference-visible geometry without updating visibility. We instead render CAD-induced motion at intermediate poses, updating visible surfaces at each step and combining the resulting motion fields.

2. Pose-dependent event support and contrast normalization. Conventional CMax typically evaluates motion hypotheses using a fixed set of events. In object tracking, however, the object occupies only part of the ROI, and its projected region changes with the unknown motion. Each candidate pose trajectory determines both how events are warped and which events are likely to belong to the object. However, the contrast objective may favor incorrect poses that shift the projected object region toward areas of high background event density, increasing the contrast score by including more events rather than correctly aligning object events. We address this with differentiable geometric support, which weights events according to the projected object region, and we propose a normalized contrast objective, which reduces the influence of variations in accumulated event weight and encourages better event alignment. Together, these components extend CMax beyond motion interpolation by incorporating candidate-dependent object visibility and event support directly into the optimization objective.

In summary, applying CMax to 6DoF object tracking introduces a different optimization problem, in which event selection and motion estimation are coupled through the unknown pose trajectory. Addressing this challenge requires a new formulation that accounts for both event alignment and trajectory-dependent event support. To demonstrate the contribution of our formulation, we first implement a baseline that directly applies conventional CMax to CAD-based object tracking, then progressively introduce our proposed components. The corresponding controlled ablations are presented in the following response.

[1] Gallego, G., Rebecq, H., & Scaramuzza, D. (2018). A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. CVPR. [2] Kang et al., Direct 3D Model-Based Object Tracking with Event Camera by Motion Interpolation. ICRA, pp. 2645–2651, 2024.

AuthorsOpenReview ↗

Controlled Ablation Study (W2/Q2)

Controlled comparisons (W2/Q2)

We construct a direct CAD-based adaptation of conventional CMax and progressively introduce the proposed components, keeping the temporal window and optimization settings fixed (this study is added in the revision, Sec. 5.4, Table 4):

  • CAD-CMax baseline: Computes constant optical flow from object velocity and CAD depth at the initial pose; we use the conventional contrast objective, IWE variance. Only events within the initial object silhouette are considered for warping, excluding events triggered in the new object location.
  • + Motion interpolation (fixed visibility): Uses interpolated poses while retaining reference-visible CAD points, without updating surface visibility.
  • + Motion interpolation (updated visibility): Renders motion fields at intermediate poses with updated visibility, including newly visible surfaces.
  • + Geometric support and normalized contrast: Weights events using candidate-dependent geometric support and normalizes the contrast by the accumulated event weight.
ConfigurationADD ↑ADD-S ↑
CAD-CMax baseline13.7623.96
+ Interpolation (fixed visibility)13.0123.52
+ Interpolation (updated visibility)14.3124.24
+ Geometric support & normalized contrast23.7636.97

The results show that interpolation alone provides no consistent improvement over the CAD-CMax baseline. Updating visibility yields only a modest gain, whereas introducing geometric support and normalized contrast improves ADD by 9.45 and ADD-S by 12.73 points over the preceding configuration. These results demonstrate that the improvement cannot be attributed simply to combining CMax with motion interpolation. Instead, the largest gains arise from accounting for trajectory-dependent event support within the contrast objective, supporting the importance of the proposed formulation. Additional ablations isolate the effects of support and normalization (Appendix A.3.2, Fig. 3), while motion-field experiments compare fused and iterative warping (Appendix A.3.1, Tables 7–8).

Presentation (W3)

We agree that the layout was compressed. We corrected overlapping elements in Figure 2 and improved spacing around figures, tables, and captions in the revised manuscript.

Thank you for the clarification. Your response helps distinguish the implementation details from the existing CMax and BIAM frameworks. However, I still view the main contributions—visibility updates during motion interpolation and candidate-dependent differentiable event support—as incremental extensions of existing mechanisms rather than a substantial methodological advance. In particular, the core ingredients of motion interpolation, object-region event selection, and normalized contrast are already established. Therefore, my concern regarding the limited methodological novelty remains.

Reviewer_ygDhOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper present a method for 6DOF tracking of an object given an event stream, or an event stream and a depth information. The pose is established by direct optimization of a cost function that measures the alignment of the events and projected features of the CAD model of the tracked object. The experimental validation shows good performance.

Critical strengths

Well presented. Nothing very novel, but a well designed method with good performance.

Critical weaknesses

The method requires an initial pose of the object. This might be very limiting in applications and in any non-lab deployment in general. Moroever, no method for pose initialization is presented.

The need for a CAD model is a significant limitation.

It seems that the depth information is available at frame rate, which will kill all the advantages of event cameras over RGB.

Questions for authors

The method, as presented needs the t0,t1 interval. How will a user set it? Since the stream of events is continuous, it seems like a needless artifact of the method? The interval is given and fixed for the whole experiment, but it should it not be adaptive, e.g. maintaining a constant number of events in the pose estimation (in case nothing is moving, there will be just noise observed in the t0,t1 interval, in the case of fast motion, event buffers might be saturated.

Important. Have any o the experiments conducted with a non-static camera? The static camera scenario is much simpler than a moving one, if this is the case, the limitation should be stated everywhere, possible even in the title (tracking from a static camera)

1 comment
AuthorsOpenReview ↗

Pose Initialization, Depth Requirements, Adaptive Windows, and Camera Motion (W1–W3, Q1–Q2)

We thank the reviewer for the constructive feedback and for recognizing the clarity and effectiveness of our method. We address the concerns below.

1. Pose Initialization and CAD Requirements (W1–W2)

Our work addresses model-based 6DoF object pose tracking, where the object's CAD model and initial pose are assumed known. This is the standard problem setting adopted by the tracking baselines in our comparison. Our contribution focuses on accurately propagating the pose from events, rather than object detection or initial pose estimation.

2. Depth Requirements and Event-Camera Advantages (W3)

Importantly, our method does not require depth for continuous tracking. In the event+depth setting, the tracker performs event-only pose updates between depth keyframes and incorporates depth-based correction only when depth becomes available. Therefore, the depth frame rate does not restrict the temporal resolution of event-based pose propagation. Our method can also operate entirely without depth, as demonstrated by our event-only experiments (Table 2; Table 3).

3. Adaptive Temporal Window Selection (Q1)

We thank the reviewer for highlighting this important challenge. We agree that selecting an appropriate temporal window is critical, particularly for our contrast-based formulation. A short window may contain insufficient object motion to produce a reliable contrast signal, while a longer window provides more motion but may violate the constant SE(3) velocity assumption (Sec. 4.2). The appropriate duration depends on object velocity, scale, and distance from the camera, making a fixed window duration or a fixed event count difficult to tune across different objects and motion conditions. To address this, our method already employs an adaptive temporal window (Sec. 4.2). We initialize it with a minimum duration (8.33 ms on Event6D, corresponding to its 120 Hz evaluation rate) and adapt it during tracking based on the history of estimated object motion. Specifically, we use previously estimated pose-induced image-plane motion to adjust the window duration, while applying a safety margin to account for estimation errors. Unlike event-count-based selection, our strategy adapts to pose-induced image motion rather than event density, reducing dependence on object texture and scale. We evaluated our adaptive window against a fixed window. We also evaluated an existing event-density-based selection strategy [1] (Table 5).

Window strategyADD ↑ADD-S ↑
Fixed window14.2841.82
Event density (d=0.15)8.2425.22
Proposed25.8454.40

Our motion-adaptive strategy achieves the best performance, while density-based adaptation performs even worse than a fixed window (Table 5). This suggests that event density alone does not reliably indicate observable object motion: a slowly moving textured object may generate more events than a rapidly moving low-texture object, making it difficult to select a suitable event-count or density threshold. Instead, our strategy relies on object geometry and motion history.

4. Tracking Under Camera Motion (Q2)

Our method estimates the object pose relative to the camera and does not require a static camera. In addition to Event6D (static camera, moving objects), we also evaluate on SPADES (Sec. 5.1), which includes trajectories where both the camera and object are moving. Our method outperforms the competing baselines on SPADES (Table 3), demonstrating its effectiveness under joint camera–object motion.

[1] Kang et al., “Direct 3D Model-Based Object Tracking with Event Camera by Motion Interpolation,” IEEE International Conference on Robotics and Automation (ICRA), pp. 2645–2651, 2024.

Reviewer_GNTPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes an event-based 6DoF tracker for objects with known CAD models. For a candidate SE(3) motion, it renders the motion field of the CAD model, warps the events with it, and optimizes the pose by maximizing the contrast of the warped events, with an adaptive time window and an optional depth term. On Event6D, it reaches 58.58/40.58 ADD-S/ADD in the event-only setting (vs. 25.11/11.44 for EDOPT) and 54.40/25.84 with depth (vs. 52.79/25.26 for EventTrack6D), or 60.51/35.12 with FoundationPose keyframe refinement. On SPADES with depth, it obtains 87.32/86.26 on synthetic and 85.40/82.26 on real sequences.

Critical strengths

The objective requires no appearance model, image reconstruction or explicit correspondences, and it clearly outperforms appearance-based event trackers in the event-only setting (Table 2: 58.58/40.58 ADD-S/ADD vs. 25.11/11.44 for EDOPT, improving on all 14 objects), with especially large gains on low-texture objects. The controlled comparison with EventTrack6D, which shares keyframe refinement and differs only in inter-frame propagation, also supports the benefit of the motion-consistency objective.

Critical weaknesses
  1. Limited novelty and ablations. Depth-induced 6DoF warps within contrast maximization (e.g., Gallego et al., 2018; Shiba et al., 2024) and direct model-based event tracking (e.g., Bryner et al., 2019; Kang et al., 2024) are established. Apart from the adaptive window, the design choices (normalized contrast, signed IWE, support weighting) are not ablated, and the closest open-source direct method (Kang et al., 2024) is not compared. Since the method has no learned component, its relevance to the ICLR audience could also be made clearer.
  2. Overstated claims. In the event+depth setting, the average gain over EventTrack6D is small (+1.6 ADD-S) and the method is notably worse on several objects (e.g., mustard, pitcher). On SPADES the depth is rendered from the CAD at the ground-truth pose, which favors direct alignment. The full tracking runtime is about 1.78 s per pose update (Appendix A.3), far from the high-speed regime that motivates the work; the faster solver substantially reduces accuracy.
Questions for authors
  1. In the event-only setting, how often is re-initialization applied and with which pose source? How do results change with the interval?
  2. Could you compare against Kang et al. (2024) and a plain CMax baseline, and ablate the main design choices?
  3. What are the end-to-end tracking rates of the proposed method and the baselines on the same hardware?
4 comments
AuthorsOpenReview ↗

CMax Novelty: Pose-Dependent Motion, Event Support, and Ablation Evidence (W1/Q2)

CMax Novelty & Ablations (W1/Q2)

We thank the reviewer and agree that our work could be better positioned relative to existing CMax methods. In particular, existing CMax formulations for camera 6DoF ego-motion estimation [1,2] address a problem closely related to ours. However, their CMax formulations do not fully address two challenges important for object tracking (Sec. 2):

Challenge 1: Accurate representation of object motion field: Existing ego-motion methods [1,2] approximate optical flow as constant over short time windows, computed from camera velocity at a fixed reference time, neglecting depth and visibility changes during motion. This causes local errors but with limited impact on global camera motion estimation, as full-image motion provides sufficient motion information and the approximation remains valid in most pixels over short windows. In object tracking, however, only a small part of the image is used and may provide insufficient motion information over such small windows, especially for small/distant objects. A larger window hurts the constant-motion approximation, requiring a better motion formulation that accounts for depth and visibility changes. 

→ We address this challenge by rendering the CAD motion at different interpolated poses to better represent the motion, and we adapt the temporal window based on the object image-induced motion (Sec. 4.1; Sec. 4.2).

Challenge 2: The events contributing to the objective. Unlike CMax formulations that use a fixed event set, the events triggered by the object are not known in advance. Their spatial support depends on the unknown pose trajectory. Each candidate motion must therefore determine not only how events are warped, but also which events should contribute to the contrast objective.

→ We address this challenge by proposing a differentiable object support region that assigns higher weights to the object events, along with a contrast function normalized by the event weights, favoring their alignment rather than increasing their number (Sec. 4.1).

Our contribution is thus a geometry-aware contrast formulation for CAD-based 6DoF object tracking, where we jointly optimize object motion and the selection of object events through a pose-dependent contrast objective. This coupling goes beyond simply applying conventional CMax to object pose tracking.

Ablation. We progressively extend a CMax baseline adapted from camera ego-motion estimation to object tracking, isolating the contribution of each component. The component comparisons are added to our manuscript in Sec. 5.4 (Table 4) and Appendix Sec. A.3.

  • CMax baseline: We compute constant optical flow from the object velocity and CAD depth rendered at the previous pose, using a fixed temporal window and IWE variance as the contrast objective. Only events within the initial object silhouette have defined motion and are warped; other events in the object bounding box remain unwarped but still contribute to the contrast score.
  • + Motion interpolation: We interpolate between the initial and estimated target poses, rendering multiple motion fields to account for changes in projected motion and visibility. All events within a fixed bounding box remain in the optimization, including those outside the warped regions.
  • + Support and normalized contrast: We introduce differentiable object support to weight events according to their geometric association with the object, and normalize the contrast objective to favor alignment rather than accumulated event weights.
ConfigurationADD↑ADD-S↑
CMax baseline13.7623.96
+ Motion interpolation14.3124.24
+ Support & normalized contrast23.7636.97

The combined gain demonstrates the importance of jointly addressing motion-field accuracy and changing spatial coverage. We added the baseline, ablations, and discussion to the revision (Sec. 5.4, Tables 4–5; Appendix Sec. A.3).

[1] Gallego, G., Rebecq, H., & Scaramuzza, D. (2018). A Unifying Contrast Maximization Framework for Event Cameras, with Applications to Motion, Depth, and Optical Flow Estimation. CVPR. [2] Shiba, S., Klose, Y., Aoki, Y., & Gallego, G. (2024). Secrets of Event-Based Optical Flow, Depth and Ego-Motion Estimation by Contrast Maximization. IEEE TPAMI, 46(12), 7742–7759.

AuthorsOpenReview ↗

BIAM Comparison, Performance, Limitations, and Re-initialization (W1/Q2, W2, Q1)

Direct Model-Based Alignment (W1/Q2)

BIAM (Kang et al., 2024) [1] relies on appearance alignment, which is sensitive to illumination, reflectance, occlusions, and model mismatches (Secs. 1–2). We instead align pose-induced event motion, avoiding appearance-based comparisons and reducing sensitivity to these variations. To demonstrate this distinction, we additionally evaluated BIAM under the same protocol. BIAM performs poorly on Event6D (7.40/1.75 vs. our 58.58/40.58 AUC ADD-S/ADD; Sec. 5.2, Table 2), where large inter-frame motion, occlusions, and event noise challenge direct alignment. On SPADES, BIAM also underperforms (Synthetic: 32.17/28.12 vs. our 67.26/56.99; Real: 8.06/5.22 vs. our 43.74/26.63 AUC ADD-S/ADD), with performance degrading at larger distances (Table 3). The full comparison is included in the revised manuscript (Tables 2 and 3).

[1] Kang et al., “Direct 3D Model-Based Object Tracking with Event Camera by Motion Interpolation,” IEEE International Conference on Robotics and Automation (ICRA), pp. 2645–2651, 2024.

ICLR Relevance (W1)

Our contributions to computer vision, robotics, and optimization align with ICLR's scope.

Performance & Limitations (W2)

Event + depth. ConTraPosE achieves the best average performance among all compared methods, including EventTrack6D, in the event+depth setting (Sec. 5.2, Table 1). We acknowledge that the improvement over EventTrack6D is small (+1.6 ADD-S points) when using only our depth-alignment correction. However, EventTrack6D relies on FoundationPose (Wen et al., 2024) for pose refinement at key depth frames. To ensure a fair comparison, we therefore evaluate an additional configuration of ConTraPosE using the same FoundationPose refinement strategy as EventTrack6D. Under this matched setting, ConTraPosE outperforms EventTrack6D on 10 of 14 objects, achieving gains of +7.72 ADD-S and +9.86 ADD AUC points on average (Table 1). These results provide stronger evidence of the benefits of our motion-consistency tracking formulation under comparable refinement settings.

SPADES depth. The simulated depth at ground-truth poses for the SPADES dataset does not give advantages to our method. We highlight, though, that the same simulated depth is provided to all methods in the event+depth setting, ensuring a fair comparison. Depth artifacts increase at larger distances (~12 m, Sec. 5.3), degrading other methods while ours maintains strong performance, suggesting that our method relies primarily on the proposed event-only tracking formulation (Table 3). 

Runtime. We acknowledge that our implementation is not real-time, as our primary contribution is a new CMax formulation. A single loss evaluation takes only 8.6 ms (Table 6); the main cost comes from the iterative PyTorch Adam solver, not the loss. To demonstrate faster inference, we implemented a batched discrete solver, reducing pose-update time from 1.78 s to 0.09 s (~20× faster). Despite lower accuracy than Adam, it achieves 34.36/17.21 ADD-S/ADD on Event6D, outperforming EDOPT, ETAP+PnP, and BIAM in the event-only setting. None of the compared baselines achieve real-time performance. We provide a runtime comparison of all methods, as suggested in the revision (Appendix Sec. A.5, Table 11).

Event-only Re-initialization (Q1)

All methods evaluated in the Event6D event-only setting use ~56 ground-truth re-detections per sequence (Sec. 5.1; Appendix Sec. A.4). Our new appendix analysis compares different re-detection rates from 0 to 56 re-detections. Higher rates limit drift for short-term evaluation; lower rates test long-term robustness. Our method is the most robust across all rates.

Reviewer_GNTPOpenReview ↗

Reply to authors

The revision and the response have addressed most my concerns. One remaining questions: the new ablation ends at 36.97/23.76 ADD-S/ADD, far below the full method (58.58/40.58). What are the differences between these two settings?

AuthorsOpenReview ↗

Clarification of Ablation Settings and Performance Differences with Table 2

We thank the reviewer for the careful follow-up and are glad the revision addressed most of the concerns. We apologize for not stating the setting of this ablation clearly in our previous response. The gap between the last row of the new ablation (36.97/23.76) and the full method in Table 2 (58.58/40.58) comes from two differences.

  • (i) Adaptive window. The new ablations test the components of the contrast-maximization (CMax) objective itself: interpolation, support weighting, normalization, and the CMax baseline. To measure their effect on the event-driven objective alone, we ran them in the event-only setting of Event6D with a fixed event window. We left out the adaptive window because it is a tracking-level component. We already ablated it in Table 5, which was in the paper before the rebuttal. That ablation used the full Event6D test set in the events + depth setting. For completeness, we have now added the adaptive window to the new ablation (event-only setting). With it, the result rises from 36.97/23.76 to 47.12/34.26 (last row of the table below). This row is the complete event-only method. It still differs from the 58.58/40.58 reported in Table 2 because it was evaluated under the reduced protocol described in (ii).
ConfigurationADD↑ADD-S↑
CMax13.7623.96
+ Motion interpolation14.3124.24
+ Support & normalized contrast23.7636.97
+ Adaptive temporal window (Full method)34.2647.12
  • (ii) Reduced evaluation protocol. The remaining gap to Table 2 comes from the evaluation protocol, which we had to shorten because of the rebuttal deadline: (a) Object subset. We used 4 representative objects (drill, scrub, spatula, bowl), chosen to cover different properties: textured vs. textureless, symmetric vs. asymmetric, and thin vs. non-thin. (b) Trajectory length. For each object, we tracked 100 consecutive steps. Because of these differences, the absolute numbers in the new table cannot be compared directly with Table 2. The table is meant to compare the components with each other. Every row, including the complete method, was run under the same protocol. In the updated paper, we describe this protocol in the table caption and the text. For the final version, we will report the ablation on the full Event6D test set (All objects, full length).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Response to All Reviewers — Summary of Revisions

We thank all three reviewers for their careful evaluation and constructive feedback. We have uploaded a revised manuscript addressing their concerns. The main changes are summarized below, with detailed explanations provided in our individual responses. All section, figure, and table references refer to the revised manuscript.

New Experiments

  • [GNTP] BIAM baseline comparison (Tables 2–3): Added direct comparisons with BIAM (Kang et al., 2024) on Event6D and SPADES to evaluate our motion-consistency formulation against appearance-based direct alignment.
  • [GNTP, 2m6L] Method component ablations (Table 4): Introduced a direct CAD-based adaptation of conventional CMax and progressively evaluated motion interpolation with fixed and updated visibility, geometric support, and normalized contrast.
  • [GNTP, 2m6L] Motion-field and contrast-objective ablations (Appendix A.3, Tables 7–8, Fig. 3): Compared fused and iterative event warping with different interpolation steps, and separately evaluated geometric support and contrast normalization.
  • [ygDh] Adaptive temporal window comparison (Sec. 5.4, Table 5): Added an event-density-based strategy to the existing fixed-window comparison, demonstrating the advantage of adapting to object-induced motion.
  • [GNTP] Re-detection sensitivity (Appendix A.4, Fig. 4): Evaluated event-only tracking across different ground-truth re-detection rates, including tracking without re-detection, to assess long-term robustness.
  • [GNTP] Runtime and optimization analysis (Table 6; Appendix A.5, Tables 9–11, Fig. 5): Added detailed runtime comparisons with baselines.

Manuscript and Presentation Changes

  • [GNTP, 2m6L] Methodological novelty (Sec. 2): Clarified the challenges of adapting conventional CMax to CAD-based object tracking, particularly changing motion fields, visibility, and object-event association.
  • [GNTP, 2m6L] Design motivations (Sec. 4.1): Added explanations motivating intermediate-pose motion rendering, geometric support, and the normalized contrast objective, linking these choices to the new ablations.
  • [GNTP, ygDh, 2m6L] Experimental discussion (Secs. 5.1–5.4; Appendix A.3–A.5): Expanded baseline descriptions, experimental settings, and discussions of the new comparisons and their main findings.
  • [2m6L] Presentation: Corrected overlapping elements in Fig. 2 and improved spacing around figures, tables, and captions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.