Pith. sign in

REVIEW 3 major objections 5 minor 61 references

Occlusion-Aware Temporally Consistent Amodal Completion for 3D Human-Object Interaction Reconstruction

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that temporally consistent amodal completion, driven by optical-flow-warped neighboring latents and template-free occlusion masks, lets monocular video of human–object interaction be reconstructed and animated in 3D.

desk verdict Temporal amodal completion is a plausible incremental contribution, but the 3D reconstruction claim rests on oracle masks and poses, so the paper oversells its end-to-end result. read the letter →

arxiv 2507.08137 v3 pith:CA26DKWX submitted 2025-07-10 cs.CV cs.AI

Hyungjun Doh , Dong In Lee , Seunggeun Chi , Pin-Hao Huang , Kwonjoon Lee , Sangpil Kim , Karthik Ramani This is my paper · ORCID
classification cs.CVcs.AI
keywords human-objectinteractionamodalcompletiontemporalconsistencymonocularvideo3DGaussiansplattingopticalflowocclusionidentificationdiffusioninpainting
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that occluded regions in monocular video of human–object interactions can be completed amodally with temporal consistency, and that the completed frames are good enough to train a 3D Gaussian Splatting representation. If the claim holds, a single video replaces the multi-view, full-visibility setups that current dynamic reconstruction methods require. The authors report that their temporal fusion and template-free mask pipeline beats prior frame-wise and video baselines on both amodal completion and temporal consistency metrics on BEHAVE and InterCap, and that the resulting 3DGS models render and animate the interactions. The strongest stated claim is that this is the first approach to photo-realistic, animatable 3D human-object interaction reconstruction from monocular video.

What carries the argument

The load-bearing mechanism is the pair of temporal modules: Bidirectional Temporal Feature Warping, which uses an off-the-shelf optical-flow network to align latent features from frames $t-n$ through $t+n$ (with $n = 7$) into frame $t$'s coordinate space, and Temporal Fusion Attention, which treats the current frame's latent as the query and the warped neighbors as keys and values in scaled dot-product attention. The machinery converts a per-frame inpainting problem into one where the diffusion model is told, in latent space, what the neighboring frames saw behind the occluder. The template-free occlusion identification supports this by supplying the mask $M_{\mathrm{occlusion}} = (M_{\mathrm{visible}} \cup \mathrm{ConcaveHull}(\pi(\mathcal{C}))) \cap M_{\mathrm{occluder}}$, where $\mathcal{C}$ is the 2D projection of the inferred 3D point cloud.

What would settle it

A reader could take a BEHAVE sequence with known ground-truth meshes and compare the predicted occlusion mask to the true occluded region computed from the meshes; if the mask IoU is low for interactions with nonconvex objects or novel poses, the temporal inpainting can only fix what the mask reveals. The end-to-end claim would also be tested by replacing the provided 6-DoF object poses used in object 3DGS training with the pipeline's own predictions.

Watch

Extended reading notes

Core claim

The paper argues that the missing piece in occlusion-heavy human-object reconstruction is temporal context, not better single-frame inpainting. Its pipeline encodes each frame with a diffusion VAE, warps the encoded features of up to fourteen neighboring frames into the current frame using optical flow, and lets the current latent attend over those warped neighbors so the inpainting model sees what the occluder hid. Occlusion masks come from projecting a template-free 3D point cloud of the interaction onto the image, forming a concave hull, and intersecting that union of visible and projected masks with the occluder's SAM2 mask. With these inputs, a Stable Diffusion inpainting pass fills the occluded regions, and the temporally coherent output frames supervise 3DGS object and human models. The authors report consistent superiority on IoU and CLIP for completion, lower warping error for temporal consistency, and higher masked PSNR, SSIM, and LPIPS for reconstruction.

Load-bearing premise

The whole pipeline rests on the assumption that a 3D model of the interaction, estimated from still images, projects onto the video frames in a way that pinpoints exactly which pixels are occluded; the paper itself allows that this estimate can be wrong for parts it has never seen.

Editorial extensions

If this is right

  • If correct, monocular RGB video is a sufficient input for photo-realistic, animatable HOI reconstruction, removing the need for multi-view rigs or full visibility.
  • Temporally consistent inpainting raises the appearance-supervision quality for 3DGS enough that masked PSNR, SSIM, and LPIPS improve over no-completion and frame-wise inpainting baselines.
  • Larger temporal windows ($n=7$) improve both completion and consistency, implying that aggregating many frames pays off when motion is complex.
  • Template-free mask identification outperforms fixed human-mask masking, so occlusion localization should be derived from 3D projection rather than predefined templates.
  • The method's gains on both BEHAVE and InterCap suggest the pipeline transfers across datasets, human subjects, and object categories.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's 'first approach' claim is about the full pipeline, but the paper's own supplementary section C.3 says object 3DGS training uses provided masks aligned with 6-DoF poses rather than predicted masks; a fully automatic monocular pipeline therefore remains to be demonstrated.
  • The temporal fusion mechanism is not HOI-specific: it could be lifted to any video inpainting or novel-view task where occlusion masks are available, so its value could be isolated in a controlled synthetic benchmark.
  • The 15–70% occlusion filter means the method was not tested on near-total occlusion, where the geometry estimate and the diffusion prior would both be stressed.
  • A testable extension is to replace the static-image HDM projection with a video-level segmentation or tracking of the occludee, which might reduce mask errors that the ablation table's human-mask baseline already hints at.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a pipeline for temporally consistent amodal completion of occluded human-object interactions in monocular video. It combines bidirectional optical-flow warping and cross-frame attention over VAE latents, a template-free occlusion mask obtained by projecting HDM 3D point clouds and intersecting them with SAM2 occluder masks, and Stable Diffusion inpainting conditioned on the fused latent. The completed frames are then used to train 3D Gaussian Splatting models (GaussianAvatar for humans and a GS-Pose-style object model) for animatable reconstruction. Experiments on BEHAVE and InterCap report amodal completion IoU/CLIP, temporal-consistency metrics, and 3D reconstruction PSNR/SSIM/LPIPS. The central claims are consistent superiority over existing baselines and the first animatable, photo-realistic 3D HOI reconstruction from monocular video.

Significance. If the full pipeline worked end-to-end, it would be a valuable zero-shot alternative to trained video inpainting for HOI: the amodal-completion stage is training-free and built from off-the-shelf optical flow, segmentation, HDM, and Stable Diffusion, and the application to 3D Gaussian Splatting is timely. The paper is transparent about several limitations and includes ablations for the temporal and masking components. However, the evidence is not yet fully aligned with the strongest claims: the quantitative advantage over Stable Diffusion inpainting is marginal on BEHAVE, and the 3D object reconstruction uses oracle masks and poses, so the end-to-end monocular reconstruction claim is not established by the reported experiments.

major comments (3)
  1. [Section 3.6, Supplementary C.3, Table 2] The 3D reconstruction experiment does not validate the full pipeline end-to-end. Supplementary C.3 states that for object reconstruction 'we use the provided object masks aligned with 6-DoF object poses instead of our predicted masks for object reconstruction.' This bypasses the proposed template-free occlusion identification of Section 3.4 and prevents Eq. (12) from being supervised by the pipeline's own masks. Table 2 therefore supports only the weaker claim that a 3DGS model trained with oracle object masks and poses can render textures from amodally completed frames. Because the first-contribution bullet in Section 1 is precisely the end-to-end monocular HOI reconstruction claim, this is load-bearing. I request either an additional experiment in which object 3DGS training uses the predicted masks (or otherwise propagates predicted mask errors), or a substantive revision of the claimed contribution to state that object masks and 6-DoF poses are assumed available.
  2. [Section 4.3, Table 1, Section 4.1] The phrase 'consistently superior performance' overstates the results in Table 1. On BEHAVE, the IoU gain over SD Inpainting is 61.75% versus 60.81% and the CLIP score is essentially tied (27.64 versus 27.63); on InterCap the CLIP gain is modest (27.65 versus 27.23). LaMa has a higher TC score on both datasets than the proposed method, although it performs worse on IoU. In addition, the evaluation is restricted to frames with occlusion ratios between 15% and 70%, and several objects are excluded for various reasons. No error bars, per-sequence breakdowns, or significance tests are reported. I ask the authors to temper the wording and to provide per-sequence statistics or significance tests so the reader can judge whether the advantage is consistent or driven by a subset of the data.
  3. [Section 3.4, Section 5, Table 4] The template-free occlusion identification claim is not yet robustly supported. The method depends on HDM's inferred 3D geometry, and Section 5 concedes that HDM 'may not always accurately infer the geometry of unseen parts.' Table 4 compares the proposed mask only against a predefined human-mask baseline; it does not quantify the accuracy of the HDM-based projected mask or the sensitivity of the amodal completion result to mask errors. Reporting mask IoU/recall against held-out occluded ground truth, or a sensitivity analysis that perturbs the proposed occlusion mask, would directly test the load-bearing assumption that the concave-hull projection of HDM point clouds localizes occluded pixels accurately enough for amodal completion.
minor comments (5)
  1. [Eq. (1), Section 3.1] The notation F_{s→e} is not defined; please clarify the meaning of s and e or replace it with standard denoising time-step notation.
  2. [Section 3.3] Please clarify whether the Temporal Fusion Attention mechanism contains any learned parameters. The equations describe a fixed scaled dot-product attention over pretrained VAE latents; if no parameters are learned, the text should say so explicitly.
  3. [Section 4.2, Supplementary D] For reproducibility, the exact text prompts used for inpainting and for the CLIP score should be listed for each object category; the paper only refers to 'a corresponding text prompt P.'
  4. [Supplementary D, 3D reconstruction] The sentence 'Since our method reconstructs only the object or the human the background' is missing a word; it should read '...the human, excluding the background.'
  5. [References [27], Section 3.6] The text cites 'GS-Pose' but the reference title is 'GS2Pose: Two-stage 6D Object Pose Estimation Guided by Gaussian Splatting.' Please make the naming consistent.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: core amodal completion is externally validated against held-out ground truth; the 3D reconstruction claim is weakened by oracle object masks and poses, but this is an attribution gap rather than a circular step.

full rationale

The derivation chain is not circular. The amodal completion and temporal consistency claims are self-contained: BTF warping uses external SEA-RAFT optical flow (Eqs. 3-4), Temporal Fusion Attention is ordinary cross-attention (Eqs. 5-6), occlusion masks are formed by projecting external HDM point clouds and intersecting with SAM2 masks (Eqs. 7-9), and inpainting uses Stable Diffusion (Eq. 11). All are evaluated against held-out ground-truth masks and frames on BEHAVE/InterCap (Table 1); no fitted parameter is renamed as a prediction (n=7 and lambda=0.2 are fixed hyperparameters chosen by ablation). HDM, SAM2, SEA-RAFT, and Stable Diffusion are external tools, and the cited self-works (e.g., [5]) are not load-bearing. The only material gap is Supplementary C.3: 'we use the provided object masks aligend with 6-DoF object poses instead of our predicted masks for object reconstruction.' Therefore Table 2's 'Ours' validates an oracle-conditioned 3DGS trained on inpainted textures rather than the complete proposed pipeline, and the Section 1 claim of 'the first approach to reconstruct photo-realistic and animatable 3D human-object interactions from monocular videos' is stronger than what is demonstrated. Section 5's admission that HDM 'may not always accurately infer the geometry of unseen parts' is a robustness caveat, not a circular step. This is an evaluation/attribution overclaim, not a circularity: the 3D reconstruction metrics are independent of the predicted masks and do not reduce by construction to the method's own inputs.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The paper does not introduce new physical entities. Its central contribution rests on several external components (HDM, SAM2, SEA-RAFT, Stable Diffusion, 3DGS) and on evaluation choices that are not fully end-to-end; the ledger lists the free hyperparameters and domain assumptions the reported results depend on.

free parameters (5)
  • Temporal support window n = 7
    Chosen empirically from Table 6 after comparing n=1 and n=3; all reported results use n=7.
  • Diffusion guidance scale = 6.0
    Set by the authors for Stable Diffusion inpainting, not justified by theory.
  • Photometric loss balance lambda = 0.2
    Chosen to balance L1 and SSIM terms in 3DGS object training.
  • Occlusion ratio inclusion range = 15%-70%
    Evaluation filter chosen to balance task difficulty; it excludes the most occluded frames despite the paper's emphasis on severe occlusions.
  • Evaluation crop margin scale = 1.2
    Used to define tight bounding boxes for PSNR-M, SSIM-M, and LPIPS-M; affects the reported metric values.
assumptions (6)
  • domain assumption Optical flow between frames can be bilinearly scaled and used to warp VAE latent features into alignment in latent space.
    Eqs. (3)-(4) assume latent-space warping with 2D optical flow captures true 3D motion; the preliminary latent-shift experiments show translation-only alignment is not sufficient under object rotation.
  • domain assumption HDM reconstructs accurate full 3D geometry of human and object from monocular RGB for mask generation.
    Section 3.4 relies on projecting HDM point clouds; Section 5 admits HDM may not infer unseen geometry accurately.
  • domain assumption Stable Diffusion inpainting, given the mask and temporally fused latent, produces semantically and geometrically plausible occluded appearance.
    Section 3.5 and the limitation section; the paper concedes reliance on SD's training distribution and on mask precision.
  • domain assumption Each input video contains exactly one human and one object.
    Section 5 explicitly states this limitation; the pipeline and masks are defined for single HOI scenes.
  • domain assumption Provided BEHAVE object masks and 6-DoF object and camera poses are accurate enough to supervise 3DGS training.
    Supplementary C.3 uses provided masks with poses instead of predicted masks; if these are inaccurate, the reconstruction claim weakens.
  • domain assumption SAM2 segmentation of inpainted frames yields valid amodal masks for IoU evaluation.
    Evaluation protocol in Section 4.2 uses SAM2 masks as model predictions; errors in SAM2 propagate directly into the IoU numbers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Occlusion-Aware Temporally Consistent Amodal Completion for 3D Human-Object Interaction Reconstruction." pith.science (2026). https://pith.science/paper/CA26DKWX

@misc{pith2026250708137,
  author       = {Pith},
  title        = {Pith review of: Occlusion-Aware Temporally Consistent Amodal Completion for 3D Human-Object Interaction Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CA26DKWX}},
  note         = {Machine review of arXiv:2507.08137}
}
read the original abstract

We introduce a novel framework for reconstructing dynamic human-object interactions from monocular video that overcomes challenges associated with occlusions and temporal inconsistencies. Traditional 3D reconstruction methods typically assume static objects or full visibility of dynamic subjects, leading to degraded performance when these assumptions are violated-particularly in scenarios where mutual occlusions occur. To address this, our framework leverages amodal completion to infer the complete structure of partially obscured regions. Unlike conventional approaches that operate on individual frames, our method integrates temporal context, enforcing coherence across video sequences to incrementally refine and stabilize reconstructions. This template-free strategy adapts to varying conditions without relying on predefined models, significantly enhancing the recovery of intricate details in dynamic scenes. We validate our approach using 3D Gaussian Splatting on challenging monocular videos, demonstrating superior precision in handling occlusions and maintaining temporal stability compared to existing techniques.

Figures

Figures reproduced from arXiv: 2507.08137 by the authors.

Figure 1
Figure 1. (left) In human–object interaction (HOI) scenarios, occlusions frequently affect both the human and the object. (middle) We inpaint the occluded regions while preserving temporal consistency for both entities across frames. (right) Leveraging the temporally consistent image sequences, we reconstruct the human and object using a 3D Gaussian splatting representation, enabling animatable 3D HOI applications. ∗Three aut… view at source ↗
Figure 2
Figure 2. Overview of Our Framework: Given a human–object interaction (HOI) monocular video, our framework performs [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Animatable 3D Reconstruction of Human–Object [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Qualitative comparison on BEHAVE [2] (Square Table, Small Table) and InterCap [11, 12] (Skateboard). Our method [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Qualitative results of 3D Reconstruction: Our [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Fixed seed preliminary experiment for temporal consistency. Two frames are completed using four different random [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Preliminary experiment using latent shift for temporal consistency. The latent feature from the first frame is shifted [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Preliminary experiment using a background mask for occlusion identification. The mask excludes only the unoccluded [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]
Figure 9
Figure 9. Figure 9: Qualitative Ablation of Our methods. score implies better visual consistency over time, as perceptually similar frames yield higher similarity in CLIP space. Flow Warping Error captures temporal misalignment at the pixel level. We first estimate optical flow between pa…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

61 extracted references · 35 canonical work pages

  1. [1]

    K. Bellock. [n. d.]. alphashape. https://github.com/bellockk/alphashape. GitHub repository, accessed 2025-04-11

    work page 2025

  2. [2]

    Bharat Lal Bhatnagar, Xianghui Xie, Ilya A Petrov, Cristian Sminchisescu, Chris- tian Theobalt, and Gerard Pons-Moll. 2022. Behave: Dataset and method for tracking human object interactions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 15935–15946

    work page 2022

  3. [3]

    A. Chen, B. Smith, and C. Lee. 2023. Amodal 3D Shape from Partial Views. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 4567–4576

    work page 2023

  4. [4]

    Seunggeun Chi, Hyung-gun Chi, Hengbo Ma, Nakul Agarwal, Faizan Siddiqui, Karthik Ramani, and Kwonjoon Lee. 2024. M2d2m: Multi-motion generation from text with discrete diffusion models. InEuropean conference on computer vision. Springer, 18–36

    work page 2024

  5. [5]

    Seunggeun Chi, Enna Sachdeva, Pin-Hao Huang, and Kwonjoon Lee. 2025. Contact-Aware Amodal Completion for Human-Object Interaction via Multi- Regional Inpainting. arXiv:2508.00427 [cs.CV] https://arxiv.org/abs/2508.00427

    arXiv 2025

  6. [6]

    Yuren Cong, Mengmeng Xu, Christian Simon, Shoufa Chen, Jiawei Ren, Yanping Xie, Juan-Manuel Perez-Rua, Bodo Rosenhahn, Tao Xiang, and Sen He. 2023. Flatten: optical flow-guided attention for consistent text-to-video editing.arXiv preprint arXiv:2310.05922(2023)

    arXiv 2023

  7. [7]

    Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. 2023. Structure and content-guided video synthesis with diffusion models. InProceedings of the IEEE/CVF international conference on com- puter vision. 7346–7356

    work page 2023

  8. [8]

    Efros, and Xiaolong Wang

    Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. 2024. COLMAP-Free 3D Gaussian Splatting. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 20796–20805

    work page 2024

Show all 61 references
  1. [9]

    Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. 2023. Tokenflow: Consis- tent diffusion features for consistent video editing.arXiv preprint arXiv:2307.10373 (2023)

    2023 arXiv

  2. [10]

    Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Sheng- ping Zhang, and Liqiang Nie. 2024. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. InProceedings of the IEEE/CVF conference on computer vision an...

    2024

  3. [11]

    Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. 2022. InterCap: Joint markerless 3D tracking of humans and objects in interaction. In DAGM German Conference on Pattern Recognition. Springer, 281–299

    2022

  4. [12]

    Yinghao Huang, Omid Taheri, Michael J Black, and Dimitrios Tzionas. 2024. InterCap: Joint Markerless 3D Tracking of Humans and Objects in Interaction from Multi-view RGB-D Images.International Journal of Computer Vision(2024), 1–16

    2024

  5. [13]

    Ajay Jain, Matthew Tancik, and Pieter Abbeel. 2021. Putting nerf on a diet: Semantically consistent few-shot view synthesis. InProceedings of the IEEE/CVF International Conference on Computer Vision. 5885–5894

    2021

  6. [14]

    Jisoo Jeong, Jamie Menjay Lin, Fatih Porikli, and Nojun Kwak. 2022. Imposing consistency for optical flow estimation. InProceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition. 3181–3191

    2022

  7. [15]

    Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan

  8. [16]

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler, and George Drettakis

  9. [17]

    M. Kim, J. Park, and K. Lee. 2023. Monocular Differentiable Rendering for Self-Supervised 3D Amodal Masks. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 789–798

    2023

  10. [18]

    Diederik P Kingma, Max Welling, et al. 2013. Auto-encoding variational bayes

    2013

  11. [19]

    Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. 2024. Hugs: Human gaussian splats. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 505–515

    2024

  12. [20]

    Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. 2018. Learning blind video temporal consistency. InProceed- ings of the European conference on computer vision (ECCV). 170–185

    2018

  13. [21]

    Dong In Lee, Hyeongcheol Park, Jiyoung Seo, Eunbyung Park, Hyunje Park, Ha Dam Baek, Sangheon Shin, Sangmin Kim, and Sangpil Kim. 2025. Editsplat: Multi-view fusion and attention-guided optimization for view-consistent 3d scene editing with 3d gaussian splatting. InProceedings...

    2025

  14. [22]

    Inhee Lee, Byungjun Kim, and Hanbyul Joo. 2024. Guess the unseen: Dynamic 3d scene reconstruction from partial 2d glimpses. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 1062–1071

    2024

  15. [23]

    P. Li, Q. Zhang, and R. Others. 2022. Compositional Models for Amodal Layout Completion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 2345–2354

    2022

  16. [24]

    Jiaqi Lin, Zhihao Li, Xiao Tang, Jianzhuang Liu, Shiyong Liu, Jiayue Liu, Yangdi Lu, Xiaofei Wu, Songcen Xu, Youliang Yan, et al. 2024. Vastgaussian: Vast 3d gaussians for large scene reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog...

    2024

  17. [25]

    Haoyu Lu, Guoxing Yang, Nanyi Fei, Yuqi Huo, Zhiwu Lu, Ping Luo, and Mingyu Ding. 2023. Vdt: General-purpose video diffusion transformers via mask modeling. arXiv preprint arXiv:2305.13311(2023)

    2023 arXiv

  18. [26]

    Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. 2024. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 20654–20664

    2024

  19. [27]

    Jilan Mei, Junbo Li, and Cai Meng. 2024. GS2Pose: Tow-stage 6D Object Pose Estimation Guided by Gaussian Splatting.arXiv preprint arXiv:2411.03807(2024)

    2024 arXiv

  20. [28]

    Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. 2021. Nerf: Representing scenes as neural radiance fields for view synthesis.Commun. ACM65, 1 (2021), 99–106

    2021

  21. [29]

    Michal Nazarczuk, Thomas Tanay, Sibi Catley-Chandar, Richard Shaw, Radu Timofte, and Eduardo Pérez-Pellitero. 2024. AIM 2024 sparse neural rendering challenge: Dataset and benchmark.arXiv preprint arXiv:2409.15041(2024)

    2024 arXiv

  22. [30]

    Nguyen, T

    H. Nguyen, T. Davis, and X. Xu. 2022. Learning Disentangled Shape-Texture for Amodal Completion. InAdvances in Neural Information Processing Systems (NeurIPS). 1–12

    2022

  23. [31]

    Ege Ozguroglu, Ruoshi Liu, Dídac Surís, Dian Chen, Achal Dave, Pavel Tokmakov, and Carl Vondrick. 2024. pix2gestalt: Amodal segmentation by synthesizing wholes. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE Computer Society, 3931–3940

    2024

  24. [32]

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. 2022. Dreamfusion: Text-to-3d using 2d diffusion.arXiv preprint arXiv:2209.14988(2022)

    2022 arXiv

  25. [33]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

    2021

  26. [34]

    Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross Girshick, Piotr Dollár, and Christoph Feichtenh...

    2024 arXiv

  27. [35]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 10684–10695

    2022

  28. [36]

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. 2022. Photorealistic text-to-image diffusion models with deep language understanding.Advances in neural inform...

    2022

  29. [37]

    Johannes L Schonberger and Jan-Michael Frahm. 2016. Structure-from-motion revisited. InProceedings of the IEEE conference on computer vision and pattern recognition. 4104–4113. MM ’25, October 27–31, 2025, Dublin, Ireland Hyungjun Doh et al

    2016

  30. [38]

    Jingyu Shi, Rahul Jain, Seunggeun Chi, Hyungjun Doh, Hyung-gun Chi, Alexan- der J Quinn, and Karthik Ramani. 2025. CARING-AI: Towards Authoring Context- aware Augmented Reality INstruction through Generative Artificial Intelligence. InProceedings of the 2025 CHI Conference on ...

    2025

  31. [39]

    Adam Sun, Tiange Xiang, Scott Delp, Fei-Fei Li, and Ehsan Adeli. 2024. Occfusion: Rendering occluded humans with generative diffusion priors.Advances in Neural Information Processing Systems37 (2024), 92184–92209

    2024

  32. [40]

    Cheng Sun, Min Sun, and Hwann-Tzong Chen. 2022. Direct voxel grid optimiza- tion: Super-fast convergence for radiance fields reconstruction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition. 5459–5469

    2022

  33. [41]

    Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Arsenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. 2022. Resolution-robust large mask inpainting with fourier convolutions. InProceedings of the IEEE/CVF ...

    2022

  34. [42]

    Teed and J

    Z. Teed and J. Deng. 2020. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. InEuropean Conference on Computer Vision (ECCV). 402–419

    2020

  35. [43]

    Yihan Wang, Lahav Lipson, and Jia Deng. 2024. Sea-raft: Simple, efficient, accurate raft for optical flow. InEuropean Conference on Computer Vision. Springer, 36–54

    2024

  36. [44]

    Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. 2004. Image quality assessment: from error visibility to structural similarity.IEEE transactions on image processing13, 4 (2004), 600–612

    2004

  37. [45]

    J. Wu, Z. Yang, and H. Kim. 2022. Self-Supervised Amodal Reconstruction from Single Images. InEuropean Conference on Computer Vision (ECCV). 341–356

    2022

  38. [46]

    Qianyang Wu, Ye Shi, Xiaoshui Huang, Jingyi Yu, Lan Xu, and Jingya Wang. 2024. Thor: Text to human-object interaction diffusion via relation intervention.arXiv preprint arXiv:2403.11208(2024)

    2024 arXiv

  39. [47]

    Xianghui Xie, Bharat Lal Bhatnagar, Jan Eric Lenssen, and Gerard Pons-Moll

  40. [48]

    Katherine Xu, Lingzhi Zhang, and Jianbo Shi. 2024. Amodal completion via progressive mixed context diffusion. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 9099–9109

    2024

  41. [49]

    Chen Yang, Sikuang Li, Jiemin Fang, Ruofan Liang, Lingxi Xie, Xiaopeng Zhang, Wei Shen, and Qi Tian. 2024. GaussianObject: High-Quality 3D Object Reconstruc- tion from Four Views with Gaussian Splatting.ACM Transactions on Graphics (TOG)43, 6 (2024), 1–13

    2024

  42. [50]

    Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. 2023. Rerender a video: Zero-shot text-guided video-to-video translation. InSIGGRAPH Asia 2023 Conference Papers. 1–11

    2023

  43. [51]

    Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang

  44. [52]

    Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. 2024. Avid: Any-length video inpainting with diffusion model. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 7162–7172

    2024

  45. [53]

    Shangchen Zhou, Chongyi Li, Kelvin CK Chan, and Chen Change Loy. 2023. Propainter: Improving propagation and transformer for video inpainting. In Proceedings of the IEEE/CVF international conference on computer vision. 10477– 10486

    2023

  46. [54]

    Shangchen Zhou, Peiqing Yang, Jianyi Wang, Yihang Luo, and Chen Change Loy

  47. [55]

    X. Zhou, Y. Li, Z. Wang, and T. Others. 2023. Amodal Instance Segmentation with Transformers. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 1234–1243

    2023

  48. [56]

    keyboard

    Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. 2024. Fsgs: Real- time few-shot view synthesis using gaussian splatting. InEuropean conference on computer vision. Springer, 145–163. Occlusion-Aware Temporally Consistent Amodal Completion for 3D Human-Object Interaction...

    2024

  49. [59]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Upscale-a-video: Temporal-consistent diffusion model for real-world video super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2535–2545

  50. [2018]

    InProceedings of the IEEE conference on computer vision and pattern recognition

    The unreasonable effectiveness of deep features as a perceptual metric. InProceedings of the IEEE conference on computer vision and pattern recognition. 586–595

  51. [2022]

    InEuropean Conference on Computer Vision

    Neuman: Neural human radiance field from a single video. InEuropean Conference on Computer Vision. Springer, 402–418

  52. [2023]

    https://repo-sam.inria.fr/fungraph/3d- gaussian-splatting/

    3D Gaussian Splatting for Real-Time Radiance Field Rendering.ACM Transactions on Graphics42, 4 (July 2023). https://repo-sam.inria.fr/fungraph/3d- gaussian-splatting/

    2023

  53. [2024]

    InIEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Template Free Reconstruction of Human-object Interaction with Proce- dural Interaction Generation. InIEEE Conference on Computer Vision and Pattern Recognition (CVPR)

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.