acceptodds
Under review as a conference paper at ICLR 2027

PanoErase:Spherical Geometry-Aware Panoramic Video Object Removal

Abstract

Panoramic video object removal aims to remove targets and associated effects while reconstructing occluded content across the viewing sphere. However, existing perspective-based methods struggle with panoramic videos, where ERP-induced seam discontinuities and polar distortions challenge both spherical reconstruction and target-effect association, while paired panoramic removal data remain scarce. To address these challenges, we construct PANOVOR, a paired panoramic video removal dataset containing 1.53M panoramic frames spanning 26.6 hours, with audited source-clean pairs and target masks. Building on PanoVOR, we propose PANOERASE, a unified model that incorporates spherical geometry into panoramic object and effect removal. PanoErase first learns panorama-aware completion from large-scale panoramic videos using SC-RoPE, which models spherical positional relationships. It is then refined with paired removal supervision through SphericalControl, which combines target-specific local evidence with panorama-wide context for effect-aware removal while preserving unrelated content. Experiments on PanoVOR-Eval and PanoVOR-Wild demonstrate consistent improvements over existing object removal methods, particularly in seam-crossing and polar regions. The dataset, code and models will be released publicly. The anonymous project website is available at https://panoerase-anonymous.pages.dev/.

Primary area:
generative models
Keywords:
360° Video Editing, Video Generation, Video Removal

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 4 reviews, above 85% of the others.
Reviewer_Q64VOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper tackles the problem of object removal in panoramic videos. Previous approaches for video object removal have focused on perspective videos. Hence a new simulated training data, approach and evaluation dataset is proposed. There qualitative examples in demo page shows the effectiveness of their approach.

Critical strengths
  1. Novel problem. Object removal approaches have mostly looked into perspective videos. The paper extends it to previously underexplored 360° videos.
  2. New Dataset. New Training and evaluation dataset has been created to tackle this novel approach.
  3. The qualitative examples show practically better removal compared to baselines.
Critical weaknesses
  1. Limited object-removal evaluation: Since GT videos are not available for PanoVOR-Wild, along with the VLLM metrics, the authors should also consider reference-free object removal metrics TokSim [1] and ReMOVE [2].

  2. Performance across different effects: Can the authors breakdown performance across different effect categories (e.g., shadows, reflections etc.)?

  3. Limited Baselines: Since SVOR and VACE are closely related to the proposed model's initialization. They are more like ablations. The authors should compare with additional (perspective) video object-removal metrics like ROSE, Minimax-remover, Object-Wiper, Omnimatte-zero etc.

  4. Contribution of pretrained models and training data: The proposed approach builds upon pretrained VACE and SVOR models, which already benefit from extensive training. Can author add a discussion or quantify how their method improves on this.

[1] Object-WIPER: Training-Free Object and Associated Effect Removal in Videos, CVPR 2026 [2] ReMOVE: A Reference-free Metric for Object Erasure, CVPRW 2024

Questions for authors

Please refer the weakness section.

Reviewer_R3MLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary
  • This paper studies object and associated-effect removal in panoramic videos.
  • A paired synthetic dataset named PanoVOR is proposed.
  • Besides the dataset, a unified model named PANOERASE is proposed, that incorporates spherical geometry into panoramic object and effect removal.
  • On PanoVOR-Eval, PanoErase improves target-region perceptual quality, effect removal, seam continuity, and polar reconstruction over the compared methods.
  • On PanoVOR-Wild, it achieves the highest quality scores from both evaluators on sampled frames.
Critical strengths
  • The proposed dataset is valuable for the the academic community
  • The geometry-aware design is well motivated. The experiment results and SphericalControl ablations provide evidence of its effectiveness.
  • The paper provides useful and clear evaluation details.
Critical weaknesses
  • The SC-RoPE comparison is incomplete. A comparison with closely related spherical encodings, such as SpheRoPE (arXiv:2606.32033), would help establish the value of the proposed coupling.
Questions for authors

None

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

This paper studies object removal in panoramic videos stored in equirectangular projection. The goal is to remove a target together with its effects, such as shadows and reflections, and fill in the background. The authors argue that the left-right seam and the stretched poles of ERP make this harder than in perspective video. They build PanoVOR, a synthetic paired dataset in which each camera path is rendered twice, with and without the target, giving 1.53M frame pairs. On top of Wan2.1-VACE they propose PanoErase, which uses a new positional encoding called SC-RoPE to handle the seam and the poles, and a conditioning branch called SphericalControl that combines tangent views around the target with features from the full panorama. The method is evaluated on a held-out synthetic split and on 124 real videos scored by VLMs, and it gives the best results on most metrics.

Critical strengths
  1. The dataset is useful and was built in a sensible way. As far as I know, there is no earlier paired dataset for panoramic video object removal. Capturing real pairs is hard for 360 video because the camera also records the operator and people nearby, so rendering each scene with and without the target is a reasonable choice. Since the clean pass is rendered again, the shadows and reflections of the target are also gone, which is what removal training needs. The dataset also includes many cases where the target crosses the seam or sits near a pole, and these cases do not exist in perspective data.

  2. SC-RoPE is a simple and well-motivated change. The paper explains clearly why a normal rotary encoding cannot be periodic across the seam and also independent of longitude at the poles, and then fixes this by coupling two rotary pairs. The change keeps the norm of the encoding, so it can be added to a pretrained model. The polar ablation in Table 4 supports it, and the idea could be reused by other panoramic diffusion models.

Critical weaknesses
  1. It is not clear whether the improvement comes from the method or from the training data. All baselines are used with their released weights and were never trained on panoramic video, while PanoErase is trained on PanoVOR and tested on data from the same renderer. A baseline fine-tuned on PanoVOR is needed for a fair comparison. This should be cheap, since SVOR uses the same backbone and is the starting point of PanoErase, so a LoRA fine-tune would be enough. There is also no ablation of the dataset, since the model is never trained without PanoVOR. The module ablations only compare variants of the authors' own model, and some gains are small; for example, SC-RoPE improves seam PSNR by only 0.17 dB on PanoVOR-Eval. The same issue affects the motivation. Figure 2 shows the seam and shadow failures with EffectErase, which has never seen panoramic video, so it does not show that these problems remain after a model is trained on panoramic data.

  2. The resolution is too low for the intended use. All results are generated at 832×416832\times416 with 81 frames, while the dataset itself is 2048×10242048\times1024. Real 360 videos are usually 4K or higher, so it is unclear whether the method is useful in practice, and the paper does not discuss this. At this resolution the polar region covers only a few latent rows, so the polar behavior is also measured in the easiest setting. The paper mentions content creation and privacy as applications but does not describe any concrete use case.

Questions for authors
  1. Could you fine-tune SVOR on PanoVOR with the same backbone and training schedule, adding only circular padding, and report it on PanoVOR-Eval?

  2. Could you train Stage II without PanoVOR, using only ROSE and VOR, and report it on PanoVOR-Eval? Together with Q1, this would show whether the gain comes from the data or from the modules.

  3. Where does Polar50 come from? It is used in Table 4 but is not described in Section 3.2 or the appendix, and its 50 cases differ from those of PanoVOR-Eval. Is it a held-out subset of PanoVOR?

  4. Do you have any results at 1024×5121024\times512 or higher?

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

The paper presents PanoErase for removing objects and their associated effects from panoramic videos. It introduces the paired PanoVOR dataset and combines SC-RoPE for spherical positional encoding with SphericalControl for target-conditioned removal. The method improves reconstruction and effect removal on PanoVOR-Eval and receives higher frame-level quality scores on real videos.

Critical strengths
  1. PanoVOR is a useful dataset contribution. It provides paired supervision for removing objects and their effects near the ERP seam and poles, which perspective datasets do not cover (§3.2).

  2. The ablations show meaningful gains. SC-RoPE improves polar reconstruction, while SphericalControl reduces Effect MAE by 18.9% under matched training (Tables 4–5).

Critical weaknesses
  1. SC-RoPE’s benefit to the final removal model is not clearly shown. Table 4 tests it only during Stage I, so it remains unclear how much this improvement carries over to object and effect removal after Stage II.

  2. The paper does not test removal at different pitch angles. Pitch changes the positions and distortion of objects and their effects in ERP images; the polar reconstruction tests and rotated mask-extraction views do not show how the full removal model handles these changes (Table 4; Appendix A.2).

Questions for authors
  1. Under the same training settings, how does the final model perform with periodic SO(2) instead of SC-RoPE?
  2. How does removal quality change at different pitch angles when the same test videos, masks, and references are rotated together?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.