acceptodds
Withdrawn from ICLR 2027

ExPO-Shape: A Benchmark and Event-Guided Novel View Synthesis for Exposure-Degraded Object Pose Estimation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Estimating single-view pose under exposure degradation requires both semantic appearance priors and illumination-robust geometric cues, neither of which can be provided by RGB or events alone. The current bottleneck of RGB–event pose estimators lies mainly in the lack of large-scale exposure-degraded benchmarks for unseen-object generalization and in the reliance on object 3D models. To fill this gap, we build ExPO-Shape, a large-scale Exposure-degraded object POse benchmark comprising 48,145 paired RGB–event samples of 9,629 unique objects across 19 categories, rendered under randomly sampled under- and over-exposure conditions. To further facilitate research on model-free, generalizable RGB-event pose estimation, we propose EVENPose, an event-guided single-view framework for robust pose estimation under exposure shifts. EVENPose stabilizes RGB embeddings with event cues instead of depth priors, guides novel-view synthesis with event-derived structure, and enforces event-level consistency to improve structural fidelity and generalization. Extensive experiments demonstrate that EVENPose serves as a strong baseline on ExPO-Shape and transfers competitively to existing RGB–event benchmarks. Overall, ExPO-Shape serves as a scalable training benchmark for model-free, generalizable RGB-event pose estimation and enables systematic evaluation of single-view robustness to exposure degradation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Exposure correction, event camera, object pose estimation, novel view synthesis

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 4 reviews, above 27% of the others.
Reviewer_2x6xOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies 3D object rotation estimation given a reference and a corresponding query image under exposure degradation, with additional event stream. It introduces ExPO-Shape, a synthetic benchmark rendered from ShapeNetCore-v2 with 48,145 RGB-event pairs of 9,629 objects in 19 categories, where event streams are simulated from frames interpolated along the trajectories under 6 under-/over-exposure levels. It also proposes EVENPose, a framework with three components: an Event-Stabilized Embedding that fuses DINOv3 features of the reference image and three event maps via cross-attention, an event-injected U-Net for novel-view embedding synthesis, and an Event Consistency Loss that compares events induced between the reference and the decoded synthesized view with the ground-truth events. EVENPose outperforms RGB-only, event-augmented, and exposure-correction baselines on five held-out ExPO-Shape categories.

Critical strengths
  1. The benchmark substantially increases object diversity over existing RGB-event pose datasets (9,629 objects in 19 categories vs. 1-21 objects), with six controlled under-/over-exposure levels. This enables evaluation of unseen-category generalization under exposure degradations, which existing RGB-event datasets do not support.
  2. The Event Consistency Loss is simple and motivated by the event triggering mechanism. It supervises novel-view synthesis in the event domain instead of exposure-sensitive pixel space. The ablation indicates that it is the most influential component (+13.3-point ACC30 on ExPO-Shape).
Critical weaknesses
  1. The event stream may leak relative motion information from reference to query. Based on the construction of ExPO-Shape, the transition path is deterministic given reference and query pose, and the number of interpolated frames is proportional to the angular distance. Therefore, the Event Count Map C, may directly indicate Δθ\Delta \theta, and the spatial distribution of events may indicate the rotation direction. This differs from existing RGB-event pose benchmarks such as YCB-Ev and YCB-Ev SD.
  2. The synthetic data may not reflect how RGB and event camera work. In the real world, RGB under-/over-exposure is due to inappropriate camera settings, and does not directly affect event camera. However, in this work, intermediate frames are first rendered with under-/over-exposure and then converted into event streams. This may lead to event streams to inherent RGB failures, e.g.: under over-exposure, the saturated region of RGB image is almost white and produce no events, whereas a real event camera would still respond.
  3. The paper mix under-exposure and low-light scenes in some discussions within method and experiment sections. Real low-light scenes introduce photon shot noise, read noise, and quantization noise [1], and event cameras may also degrade in low light, neither of which is modeled by the noise-free exposure scaling in ExPO-Shape. Therefore, the results on ExPO-Shape may support robustness to camera exposure settings rather than to low-light conditions.

[1] Kaixuan Wei et al., Physics-based noise modeling for extreme low-light photography. IEEE TPAMI, 2022.

Questions for authors
  1. In YCB-Ev SD, each pose has its own short event sequence, and the views within a scene do not form a continuous trajectory. What event stream is used as the events from Ir to Iq when evaluating on YCB-Ev SD?
  2. Are the intermediate frames stored as 8-bit images or as linear float/HDR values, and how is each exposure level implemented? Are the events generated from the exposure-degraded frames or from clean frames?
  3. Since NOPE is the primary baseline, why is NOPE+E not reported in Table 1?
  4. In Fig. 5, the baseline on ExPO-Shape (25.2/61.8) matches RelPose in Table 1, while the baseline on YCB-Ev SD (15.4/69.1) matches NOPE. Which method is the baseline?
  5. The embedding e^\hat{e} is built on DINOv3 features with learnable adapters. Which pretrained VAE decoder maps it to an image?
Reviewer_3K2KOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces ExPO-Shape, a large-scale synthetic RGB–event benchmark for object pose estimation under severe exposure degradation, containing 48145 samples across 9629 objects and 19 categories. The authors further propose EVENPose, an event-guided framework that integrates Event-Stabilized Embedding, Event-Guided Novel View Synthesis, and Event Consistency Loss to improve pose estimation without requiring explicit 3D object models at inference time. Experiments demonstrate improvements over existing baselines on ExPO-Shape and competitive performance on existing RGB–event benchmarks.

Critical strengths
  • ExPO-Shape addresses the lack of large-scale RGB-event datasets for exposure-degraded object pose estimation.
  • The proposed framework incorporates event-derived structural cues into feature representations, novel-view synthesis, and training supervision. The Event Consistency Loss is particularly interesting, as it encourages geometrically consistent predictions.
  • EVENPose outperforms existing baselines on ExPO-Shape and demonstrates improvements on real-world YCB-Ev data. The ablation studies also support the effectiveness of the proposed components.
Critical weaknesses
  • The proposed benchmark relies entirely on synthetic RGB images and simulated events. Although the authors evaluate on real YCB-Ev data, they do not analyze the distribution gap between synthetic and real event data, such as differences in event density, noise, temporal characteristics, and spatial patterns. This raises concerns about the realism of the benchmark and the generalizability of the learned representations.
  • The reference and query images are generated under the same exposure condition. This simplifies the problem and does not reflect scenarios where exposure varies between observations.
  • The method heavily depends on event-derived geometric cues, but its robustness to event sparsity, sensor noise, and different motion conditions is not sufficiently investigated. These factors could significantly affect performance in practical applications.
Questions for authors
  • How does EVENPose perform when the reference and query images have different exposure levels?
  • Could the authors provide a quantitative or qualitative comparison of the distributions of synthetic and real event data?
Reviewer_H7nDOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces ExPO-Shape, a synthetic RGB–event benchmark for relative object rotation estimation under exposure degradation, and EVENPose, which incorporates event features into a novel-view embedding synthesis framework. The method combines event-based feature fusion, event-guided synthesis, and an event consistency loss

Critical strengths

A new task formulation with an associated benchmark and proposed solutions.

Critical weaknesses
  • The physical justification of the central event consistency loss is not sufficiently established. In Eqs. (6)–(10), the loss selects a polarity according to the net event count, then compares the endpoint log-intensity difference against the total count of that polarity. However, under the ideal event model, the endpoint difference approximately constrains the threshold-weighted signed count, Δlog⁡I≈C+N+−C−N−,\Delta\log I \approx C_+N_+-C_-N_-, rather than either polarity’s total count. For pixels containing both positive and negative events, selecting one polarity does not account for cancellation. Consequently, even physically correct endpoint predictions can incur a nonzero penalty beyond the usual threshold residual. This does not invalidate the empirical usefulness of ECL, but it weakens its interpretation as a physically consistent constraint. Clarifying the underlying approximation and comparing it with a signed-count formulation would directly address this concern.

  • The methodological gain beyond event fusion could be isolated more convincingly. The ESE-augmented PIZZA/RelPose baselines and the component ablations are appreciated. However, they do not clearly establish how much improvement the full framework provides over a matched NVS baseline with straightforward event fusion. A single controlled comparison—using the same feature backbone, adapters, and training protocol, but without EG-NVS or ECL—would clarify whether the additional machinery provides a substantial benefit beyond incorporating event observations. This is particularly relevant because the method builds on an existing NVS-based pose estimation framework

Overall, the benchmark is potentially useful, and the reported improvements are encouraging. My reservations concern the justification of the central loss and the evidence establishing the method’s contribution beyond event fusion, rather than the relevance of the task. These issues currently prevent me from recommending acceptance.

Questions for authors

See Above

Reviewer_4fETOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies RGB-event pose estimation for unseen objects under exposure degradation. It introduces ExPO-Shape, comprising 48,145 paired samples from 9,629 objects across 19 categories, and proposes EVENPose. Given a reference RGB image, a query RGB image, and the intervening event stream, the method incorporates Event-Stabilized Embedding (ESE), Event-Guided Novel View Synthesis (EG-NVS), and an Event Consistency Loss (ECL) into a framework based on candidate-view feature synthesis and matching. The described formulation and evaluation primarily concern relative rotation. Experiments cover ExPO-Shape, YCB-Ev SD, and real YCB-Ev, together with component ablations, efficiency measurements, and failure analysis.

Critical strengths
  1. The problem is practically relevant, and the benchmark has potential value.
    Exposure degradation compromises appearance cues used for object pose estimation. ExPO-Shape offers substantial object and category diversity, with controlled exposure conditions and a category-based split that supports studying exposure robustness and generalization to unseen categories. Rendering intermediate trajectories and densely interpolating frames also accounts for the motion process more explicitly than event simulation based only on two endpoint images.
  2. The complete system reports substantial gains and includes preliminary real-data validation.
    On ExPO-Shape, the mean ACC 30 reaches 52.8, compared with 31.0 for the strongest listed baseline, CoTF+NOPE. On real YCB-Ev, it reaches 24.7, compared with 16.5 for the strongest listed RGB baseline. The component ablations also provide preliminary evidence of conditional contributions within the full architecture. These results support the feasibility of the research direction.
Critical weaknesses
  1. The endpoint-consistency interpretation of ECL lacks necessary assumptions.

    Section 3.3, Equations (6)–(10), estimates event counts from the log-intensity difference between endpoint images and compares these estimates with accumulated positive and negative event counts separately. However, endpoint intensity differences generally reflect the net change induced by events, rather than the accumulated count of either polarity.

With equal positive and negative threshold magnitudes cc, no noise, and negligible quantization residual, the correct endpoint satisfies: ΔL/c≈N+−N−.\Delta L/c \approx N_+ - N_-. If a pixel produces N+=5N_+=5 and N−=4N_-=4, the correct endpoint has ΔL/c≈1\Delta L/c \approx 1. Under the paper’s positive mask and loss definition, the unweighted positive per-pixel term is approximately (1−5)2=16(1-5)^2=16, while it is minimized at ΔL/c=5\Delta L/c=5. Thus, the concern is not merely a nonzero penalty at the correct endpoint: the loss may favor an incorrect endpoint intensity.

This counterexample does not establish how prevalent mixed-polarity pixels are in the dataset or imply that joint training necessarily fails. Nevertheless, the paper should specify the count definitions, temporal windows, and preprocessing, establish conditions under which this bias is avoided, and quantify its practical impact. Ablation improvements alone do not validate the proposed physical interpretation.

  1. The experiments do not yet adequately separate the contributions of different changes, and some result configurations are unclear.

    Assuming other settings are held fixed, the component ablations in Figure 5 support conditional contributions within the full model. However, original NOPE uses pretrained VAE features, whereas EVENPose uses DINOv3, and the manuscript does not clearly establish whether the NOPE baseline uses matched encoders and adapters. A clearly specified encoder-matched RGB baseline and a simple event-fusion baseline are therefore needed to distinguish improvements from feature representations, additional event observations, and the proposed event-processing designs.

    Moreover, Table 2(b) on real YCB-Ev contains only RGB-only baselines and does not include event-augmented baselines. It supports the complete system’s advantage over the listed RGB methods, but provides limited evidence of superiority over methods receiving the same modalities.

    The ablation results also require reconciliation. NOPE in Table 1 reports 25.6 / 58.0 for ACC 30 / Median, whereas the Baseline in Figure 5 reports 25.2 / 61.8. The full model in Figure 5 reports 52.8 / 26.1, whereas L2+ECL in Figure 6 reports 49.1 / 29.6. Different settings, runs, or labeling could explain these differences, but the dataset and configuration mappings are currently insufficiently documented for independent verification.

Questions for authors
  1. Are N+N_+ and N−N_- the raw event counts accumulated over the entire reference-to-query interval? How are pixels containing both polarities handled? Please report their proportion among event-active pixels and explain whether the ECL target remains consistent with the correct endpoint in this case.

  2. How are the features produced by DINOv3 and the adapters interfaced with the pretrained VAE decoder? Please specify representation dimensions, any latent-space mapping or alignment mechanism, and which components are frozen or trained.

  3. Do the baselines use matched encoders, adapters, training conditions, and pose-search settings? Are results available for simple event fusion with the same backbone and for event-augmented baselines on real YCB-Ev?

  4. Which model does the Baseline in Figure 5 represent? Which dataset and configuration are used in Figure 6? Please explain their differences from the main tables and other ablation results.

  5. Does the method estimate only relative rotation? If translation is also estimated, please provide the method and evaluation. Otherwise, please clarify the scope and explain the sources of translation and scale used in the qualitative 3D-box visualizations.

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

Related papers

Discussion (0)

Sign in to comment.