SRHand: Super-Resolving Hand Images and 3D Shapes via View/Pose-aware Neural Image Representations and Explicit Meshes

Common Response to Reviewers

We sincerely thank the reviewers for their constructive feedback. We appreciate all reviewers recognizing the motivation of our work and the novelty of the method, along with the SOTA results. We have conducted several additional experiments to address remaining concerns.

1. 3D reconstruction experiment with other hand reconstruction methods: HARP[R1] , LiveHand [R2].

2. 3D Consistency-considered super-resolution experiment including Gaussian splatting, SRGS [R3].

3. Extreme resolution experiment to show the lower-bound of SRHand, and the results for different occlusion rates.

Please refer to the additional experiment results for your consideration.

References:
[R1] Korrawe Karunratanakul, Sergey Prokudin, Otmar Hilliges, and Siyu Tang. Harp: Personalized hand reconstruction from a monocular rgb video. In CVPR, 2023.

[R2] Akshay Mundra, Mallikarjun B R, Jiayi Wang, Marc Habermann, Christian Theobalt, and Mohamed Elgharib. Livehand: Real-time and photorealistic neural hand rendering. In ICCV, 2023.

[R3] Xiang Feng, Yongbo He, Yan Yang, and Yubo Wang, Yifei Chen, Zhan Wang, Zhenzhong Kuang, Feiwei Qin, Jiajun Ding. SRGS: Super-Resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024.

Q1: Ambiguity in Terminology.
A1: We had included 'known_typo_main_figure.pdf' in the submitted supplementary file mentioning those typos.

Q2: Innovation about GIIF module.
A2: The main contribution of SRHand lies not in GIIF alone, but in the joint optimization of the SR module and the 3D reconstruction module, which enables consistent super-resolved images and high-resolution animatable hand meshes to be learned simultaneously. GIIF supports this pipeline as GIIF key modules are used in the joint optimization and affect the 3D reconstruction stage. Integrating GIIF into a tightly coupled SR–3D pipeline for dynamic, low-resolution hand avatars, to the best of our knowledge, is a novel contribution. This differs fundamentally from prior image SR methods, which operate independently of 3D learning or induced by 3D consistency for static objects or scenes.

Replacing GIIF with LIIF (or LIIF* variants) leads to a noticeable drop in both SR quality (Tab. 1) and 3D reconstruction accuracy (Tab. 3). Importantly, the adaptive fine-tuning stage, where GIIF and the mesh module are updated together, reduces P2P error from 3.09mm to 2.16mm, highlighting that our novelty stems from the SR–3D joint optimization and module sharing across stages.

Thank you again for your constructive feedback. We have provided a rebuttal to address your concerns, and we hope that our answers have sufficiently addressed them. If there are any remaining points you would like us to clarify, we would greatly appreciate your further feedback and would be glad to discuss them.

Common Response to Reviewers

We sincerely thank the reviewers for their constructive feedback. We appreciate all reviewers recognizing the motivation of our work and the novelty of the method, along with the SOTA results. We have conducted several additional experiments to address remaining concerns.

1. 3D reconstruction experiment with other hand reconstruction methods: HARP[R1] , LiveHand [R2].

2. 3D Consistency-considered super-resolution experiment including Gaussian splatting, SRGS [R3].

3. Extreme resolution experiment to show the lower-bound of SRHand, and the results for different occlusion rates.

Please refer to the additional experiment results for your consideration.

References:
[R1] Korrawe Karunratanakul, Sergey Prokudin, Otmar Hilliges, and Siyu Tang. Harp: Personalized hand reconstruction from a monocular rgb video. In CVPR, 2023.

[R2] Akshay Mundra, Mallikarjun B R, Jiayi Wang, Marc Habermann, Christian Theobalt, and Mohamed Elgharib. Livehand: Real-time and photorealistic neural hand rendering. In ICCV, 2023.

[R3] Xiang Feng, Yongbo He, Yan Yang, and Yubo Wang, Yifei Chen, Zhan Wang, Zhenzhong Kuang, Feiwei Qin, Jiajun Ding. SRGS: Super-Resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024.

Q1: Experiments on extremely small image resolution and complex occlusions.
A1: Our main experiments, ×16 upscale factor, already target a realistic setting in both human and hand reconstruction literature, where the hand occupies 1.5% (L24) of the full body size. To further probe the lower limit of SRHand, we conducted an additional experiment with an upscaling factor of ×32 (where the input images are as small as 8×8). While this factor goes beyond typical practical scenarios, it serves as a stress test to evaluate how well the joint SR–3D pipeline preserves geometry and texture. Thanks to the strong combination of the SR module and the 3D reconstruction module, the joint optimization maintains comparable performance even under this extreme setting, as shown in the table below. Furthermore, seeing the arbitrary input resolution experiments presented in Supplementary Section 3.5 together, these results demonstrate that SRHand scales consistently: as the input image resolution increases, 3D reconstruction quality improves accordingly.

Table 1: Quantitative results of the arbitrary scale input image experiment, including an extreme resolution image.

Q2: Validation under complex occlusions.
A2: We also present complex occlusion experiments. Hand reconstruction is well-known to be challenging due to frequent self-occlusions and highly articulated poses. Our multi-view video setup is designed to alleviate these issues, yet detailed geometric reconstruction remains difficult in occluded regions. To simulate such challenging conditions, we reduce the number of training cameras to one or two and sample pose-varying frames. The occlusion rate is quantified through the percentage of invisible vertices of the hand in the camera viewing space. As expected, the reconstruction accuracy decreases as the occlusion rate increases, shown in the table below; however, not drastically.

Table 2: Quantitative results under diverse occlusion rate.

Q3: Missing Reference.
A3: I2UV-HandNet [1] focuses on reconstructing a high-resolution UV-map hand mesh (without texture) from a single image, which has a different objective compared to ours. Nevertheless, its combination of super-resolution and hand geometry reconstruction shares conceptual similarities with our approach, and we will include I2UV-HandNet as a related work in the final version.

Reference:
[1] Ping Chen, Yujin Chen, Dong Yang, Fangyin Wu, Qin Li, Qingpei Xia, Yong Tan. I2UV-HandNet: Image-to-UV Prediction Network for Accurate and High-fidelity 3D Hand Mesh Modeling. In ICCV, 2021.

Common Response to Reviewers

We sincerely thank the reviewers for their constructive feedback. We appreciate all reviewers recognizing the motivation of our work and the novelty of the method, along with the SOTA results. We have conducted several additional experiments to address remaining concerns.

1. 3D reconstruction experiment with other hand reconstruction methods: HARP[R1] , LiveHand [R2].

2. 3D Consistency-considered super-resolution experiment including Gaussian splatting, SRGS [R3].

3. Extreme resolution experiment to show the lower-bound of SRHand, and the results for different occlusion rates.

Please refer to the additional experiment results for your consideration.

References:
[R1] Korrawe Karunratanakul, Sergey Prokudin, Otmar Hilliges, and Siyu Tang. Harp: Personalized hand reconstruction from a monocular rgb video. In CVPR, 2023.

[R2] Akshay Mundra, Mallikarjun B R, Jiayi Wang, Marc Habermann, Christian Theobalt, and Mohamed Elgharib. Livehand: Real-time and photorealistic neural hand rendering. In ICCV, 2023.

[R3] Xiang Feng, Yongbo He, Yan Yang, and Yubo Wang, Yifei Chen, Zhan Wang, Zhenzhong Kuang, Feiwei Qin, Jiajun Ding. SRGS: Super-Resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024.

Q1: Using GT Hand Pose.
A1: In the main paper, we used ground-truth pose parameters to provide an upper bound and to isolate the effect of our SR–3D reconstruction pipeline. This evaluation protocol is a widely adopted strategy in prior works [1, 2, 3, 4, 5], which also leverage GT SMPL parameters to reconstruct detailed shapes from incomplete, occluded, or degraded images while decoupling pose estimation errors from reconstruction quality.

To further assess the robustness of SRHand over pose errors, we conducted experiments in the supplementary Section 3.4 by adding Gaussian noise to the GT parameters. In a multi-view video capture setup, global constraints allow pose parameters to be fitted jointly across views, reducing the joint error to mm scale [6, 7]. With the perturbations up to 9.54 mm (~1cm) per joint, SRHand maintained stable performance, remaining tolerant to realistic levels of pose inaccuracy. Nonetheless, extreme pose errors limit performance, as is common in most hand reconstruction pipelines.

Regarding the question of the low-resolution settings and GT existness, in a multi-view body capture system, hand region resolution dynamically changes depending on its pose and camera viewpoint. Even in the same pose, one camera may capture a small region of the hand while others capture a larger region. For example, in the ActorsHQ [8] dataset, within the same pose, a single hand region varies (45×46 ~ 108×127) depending on its camera views. Thus, even though the camera captures the hand in a low resolution, global 3D pose parameters can still be extracted across multiple cameras.

In addition, considering this variability observed in real datasets, we evaluated SRHand on multiple input resolutions, including 48×48, which approximates the hand region resolution in real-world datasets (ActorsHQ in Table 3 for reviewer yhNh and Supplementary Section 3.5; see also Table 1 below). We find out that as the input image resolution increases, the reconstruction accuracy typically improves. We also include experiments with even smaller hand crops to probe the lower limit of SRHand. As noted in another review (sv5s), which requested testing at extreme small scales, SRHand maintained faithful reconstruction results without significant degradation compared to other resolution scales. Nevertheless, estimating accurate hand poses from extremely low-resolution images remains an open research challenge.

Table 1: Quantitative results of arbitrary scale input image experiment.

Q2: More experiments for 3D content generation.
A2: Certain works (e.g., 3DEnhancer [9] and LGM [10]) do not align with our research goal and are difficult to compare for several reasons. First, generating 3D objects is entirely different from reconstructing detailed personalized hand avatars from captured images. Reconstructing a personalized avatar is to optimize the model to capture finer details across input images, whether a generalizable 3D object generation task is to work upon a large-scale pretraining model (e.g MVDream [11]). None of the prior avatar reconstruction works [12, 13, 14] has been compared with the methods for 3D generic object generation tasks. Second, their architectures are specifically designed for static objects and scenes, not an animatable avatar. Their proposed modules, such as multi-view row attention, near-view epipolar aggregation, are designed for static objects and are not straightforward to extend for dynamic human avatars. Their training codes are also not publicly available. Regarding the existing diffusion-based models, IDM [15] and DiSR-NeRF [16] are both diffusion-based methods and are included as part of the comparative experiments in our main paper.

Instead, we include experiments with SRGS [17], which learns Gaussian splatting with super-resolved images for static scenes. As SRGS supports a fixed upscale factor, we arbitrarily control its factor through iterative upscaling. As shown below, our method achieves performance comparable to other baselines for static 3D hand reconstruction, while still supporting animatable hand avatars. Furthermore, as the upscaling factor increases, SRHand significantly outperforms the prior works, particularly in perceptual metrics such as LPIPS and SSIM. This is because PSNR measures only the mean squared pixel difference, where overly smooth or blurred textures can still yield high scores, whereas LPIPS and SSIM better capture perceptual fidelity and structural detail.

Table 2: Quantitative comparison of 3D consistency considered SR methods. NeRF-SR and DiSR-NeRF are brought from the main paper for comparisons.

Reference:
[1] Junying Wang, Jae Shin Yoon, Tuanfeng Y. Wang, Krishna Kumar Singh, Ulrich Neumann, Complete 3D human Reconstruction from a Singe Incomplete Image. In CVPR, 2023.

[2] Jihyun Lee, Minhyuk Sung, Honggyu Choi, and Tae-Kyun Kim. Im2hands: Learning attentive implicit representation of interacting two-hand shapes. In CVPR, 2023.

[3] Zeren Jiang, Chen Guo, Manuel Kaufmann, Tianjian Jiang, Julien Valentin, Otmar Hilliges, Jie Song. MultiPly: Reconstruction of Multiple People from Monocular Video in the Wild. In CVPR, 2024.

[4] Minje Kim, and Tae-Kyun Kim. BiTT: Bi-directional Texture Reconstruction of Personalized Interacting Two Hands from a Single Image. In CVPR, 2024.

[5] Kennard Yanting Chan, Fayao Liu, Guosheng Lin, Chuan Sheng Foo, and Weisi Lin. Robust-PIFU: Robust Pixel-Aligned Implicit Function for 3D Human Digitalization from A Single Image. In ICLR, 2025.

[6] Wei Xie, Zhipeng Yu, Zimeng Zhao, Binghui Zuo, Yangang Wang. HMDO : Markerless Multi-view Hand Manipulation Capture with Deformable Objects, In Graphic Models, 2023.

[7] Xiaozheng Zheng, Chao Wen, Zhou Xue, Pengfei Ren, Jingyu Wang1. HaMuCo: Hand Pose Estimation via Multiview Collaborative Self-Supervised Learning. In ICCV, 2023.

[8] Mustafa Işık, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, Matthias Nießner. HumanRF: High-Fidelity Neural Radiance Fields for Humans in Motion. In ACM TOG, 2023.

[9] Yihang Luo, Shangchen Zhou, Yushi Lan, Xingang Pan, and Chen Change Loy. 3DEnhancer: Consistent Multi-View Diffusion for 3D Enhancement. arXiv:2412.18565, 2024.

[10] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation. In ECCV, 2024.

[11] Yichun Shi, Peng Wang, Jianglong Ye, Long Mai, Kejie Li, and Xiao Yang. MVDream: Multi-view Diffusion for 3D Generation. In ICLR, 2024.

[12] Tianjian Jiang, Xu Chen, Jie Song1, and Otmar Hilliges. InstantAvatar: Learning Avatars from Monocular Video in 60 Seconds. In CVPR, 2023.

[13] Hezhen Hu, Zhiwen Fan, Tianhao Walter Wu, Yihan Xi, Seoyoung Lee, Georgios Pavlakos, and Zhangyang Wang. Expressive Gaussian Human Avatars from Monocular RGB Video. In NIPS, 2024.

[14] Hsuan-I Ho, Jie Song, and Otmar Hilliges. SiTH: Single-view Textured Human Reconstruction with Image-Conditioned Diffusion. In CVPR, 2024.

[15] Sicheng Gao, Xuhui Liu, Bohan Zeng, Sheng Xu, Yanjing Li, Xiaoyan Luo, Jianzhuang Liu, Xiantong Zhen, and Baochang Zhang. Implicit diffusion models for continuous super-resolution. In CVPR, 2023.

[18] Jie Long Lee, Chen Li, and Gim Hee Lee. Disr-nerf: Diffusion-guided view-consistent super-resolution nerf. In CVPR, 2024.

[19] Xiang Feng, Yongbo He, Yan Yang, and Yubo Wang, Yifei Chen, Zhan Wang, Zhenzhong Kuang, Feiwei Qin, Jiajun Ding. SRGS: Super-Resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024.

Please let us know when you still have more concerns. If your concerns have been sufficiently resolved, could you kindly update your rating accordingly?

Common Response to Reviewers

We sincerely thank the reviewers for their constructive feedback. We appreciate all reviewers recognizing the motivation of our work and the novelty of the method, along with the SOTA results. We have conducted several additional experiments to address remaining concerns.

1. 3D reconstruction experiment with other hand reconstruction methods: HARP[R1] , LiveHand [R2].

2. 3D Consistency-considered super-resolution experiment including Gaussian splatting, SRGS [R3].

3. Extreme resolution experiment to show the lower-bound of SRHand, and the results for different occlusion rates.

Please refer to the additional experiment results for your consideration.

References:
[R1] Korrawe Karunratanakul, and et al. Harp: Personalized hand reconstruction from a monocular rgb video. In CVPR, 2023.

[R2] Akshay Mundra, and et al. Livehand: Real-time and photorealistic neural hand rendering. In ICCV, 2023.

[R3] Xiang Feng, and et al. SRGS: Super-Resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024.

Q1: Compare with more valid baselines.
A1: For 3D hand reconstruction baselines, HARP [1], LiveHand [2], and other prior works (Handy [3], HandAvatar [4]) were not included as XHand [5] and UHM [6] significantly outperform these methods. Nevertheless, we present additional experiments below for HARP and LiveHand, which are among the recent works. HARP reconstructs a 3D hand using mesh subdivision with 3,093 vertices, which is not sufficient to represent a detailed hand geometric shape (significantly less than ours: 49,281 vertices). Also, assuming a one-point light condition differs from the benchmark settings used in our experiments, and yields geometric artifacts. LiveHand employs low-resolution NeRF combined with a super-resolution network for real-time hand rendering. While efficient, LR NeRF and CNN-based SR module likely loses the 3D geometric information. Both methods encounter difficulties in reconstructing 3D hands from inconsistent super-resolved images and lack performance compared to ours.

Table 1: Quantitative comparison of 3D reconstruction results. For fair comparison, our GIIF module was used for super-resolving images.

Mip-NeRF [7] tackles a different objective from ours; while it renders at across continuous scales, it remains bounded to the scale of training images, not performing super-resolution. Prior works of super-resolution on top of NeRF [8, 9, 10], therefore, do not directly compare with Mip-NeRF, only mentioning it as a related work.

Instead, we include experiments with SRGS [11], which learns Gaussian splatting from super-resolved images for static scenes. As SRGS supports a fixed upscale factor, we arbitrarily control its factor through iterative upscaling. As shown below, our method achieves a performance comparable to other baselines designed for static scene reconstruction, while supporting dynamic and animatable hand avatars, which the other methods cannot afford to. As the upscaling factor increases, SRHand significantly outperforms the prior works, particularly in perceptual metrics such as LPIPS and SSIM. This is because PSNR measures only the mean squared pixel difference, where overly smooth or blurred textures can still yield high scores, whereas LPIPS and SSIM better capture perceptual fidelity and structural details.

Table 2: Quantitative comparison of 3D consistency considered SR methods. NeRF-SR and DiSR-NeRF are brought from the main paper for comparisons.

Q2: Adjust the story and focus on 3d avatar creation.
A2: Our paper is presently written in the order of process (SR $\rightarrow$ 3D reconstruction $\rightarrow$ fine-tuning). The current structure, introducing the SR module first in the method section and presenting the SR experiments before 3D results, might overemphasize the SR module. To counteract this, we will refine the narrative and adjust the section flow to make the 3D hand avatar creation the clear focal point. Specifically:

1. Method structure (Sec. 3): We will revise the method section to introduce SRHand as a unified 3D hand avatar reconstruction framework. The SR module (GIIF) will be described as a sub-module to support SR–3D unified structure instead of treating them as sequentially independent tasks.

2. Experiment ordering (Sec. 4): We will reorder the experimental results to present 3D reconstruction metrics and qualitative meshes first, followed by SR evaluations as an auxiliary task. Additionally, we also plan to balance the section length to mainly focus on the 3D avatar results and discussions.

3. Introduction (Sec 1.) & Conclusion (Sec 2.): We will revise the introduction to clearly convey that the SR module is an integral part of the 3D reconstruction pipeline, designed to enable high-fidelity mesh recovery under low-resolution, pose-varying conditions rather than serving as an independent objective. The conclusion will be updated accordingly to emphasize SR–3D perspective as the central contribution of the work.

Q3: Hand Occupancy Region.
A3: The statement (L24) is based on our empirical observation across the full-body image datasets. We computed the pixel-wise ratio between the region using tight bounding boxes of hand and full-body in several representative samples. We present its comparisons in the table below, where a single hand occupies up to 1.5% of the whole depending on its pose.

Table 3: Pixel-level hand region comparison to body region across multiple datasets. (BB stands for bounding box.)

Q4: Negative Social Impact.
A4: We agree that hand geometry and motion patterns can implicitly contain biometric cues. Our model is trained on the InterHand2.6M dataset, which is collected for public research purposes. Furthermore, the GIIF learns a generalized hand prior, not identity labels, and our 3D reconstruction pipeline cannot encode personal identifiers as long as they are absent in input images. This design setting minimizes privacy risk; still, some texture information related to personal identity can be implicitly learned in the GIIF network. We will clarify this social impact in the final version.

References:
[1] Korrawe Karunratanakul, Sergey Prokudin, Otmar Hilliges, and Siyu Tang. Harp: Personalized hand reconstruction from a monocular rgb video. In CVPR, 2023.

[2] Akshay Mundra, Mallikarjun B R, Jiayi Wang, Marc Habermann, Christian Theobalt, and Mohamed Elgharib. Livehand: Real-time and photorealistic neural hand rendering. In ICCV, 2023.

[3] Rolandos Alexandros Potamias, Stylianos Ploumpis, Stylianos Moschoglou, Vasilios Triantafyllou, and Stefanos Zafeiriou. Handy: Towards a high fidelity 3D hand shape and appearance model. In CVPR, 2023.

[4] Xingyu Chen, Baoyuan Wang, and Heung-Yeung Shum. Hand avatar: Free-pose hand animation and rendering from monocular video. In CVPR, 2023.

[5] Gyeongsik Moon, Weipeng Xu, Rohan Joshi, Chenglei Wu, and Takaaki Shiratori. Authentic hand avatar from a phone scan via universal hand model. In CVPR, 2024.

[6] Qijun Gan, Zijie Zhou, and Jianke Zhu. Xhand: Real-time expressive hand avatar. arXiv:2407.21002, 2024.

[7] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-NeRF: A Multiscale Representation for Anti-Aliasing Neural Radiance Fields. In CVPR, 2021.

[8] Chen Wang, Xian Wu, Yuan-Chen Guo, Song-Hai Zhang, Yu-Wing Tai, and Shi-Min Hu. Nerf-sr: High-quality neural radiance fields using supersampling. In ACM, 2022.

[9] Xudong Huang, Wei Li, Jie Hu, Hanting Chen, and Yunhe Wang. RefSR-NeRF: Towards High Fidelity and Super Resolution View Synthesis. In CVPR, 2023.

[10] Jie Long Lee, Chen Li, and Gim Hee Lee. Disr-nerf: Diffusion-guided view-consistent super-resolution nerf. In CVPR, 2024.

[11] Xiang Feng, and et al. SRGS: Super-Resolution 3D Gaussian Splatting. arXiv:2404.10318, 2024.

[12] Julieta Martinez, and et al. Codec avatar studio: Paired human captures for complete, driveable, and generalizable avatars. In NeurIPS Track on Datasets and Benchmarks, 2024.

[13] Thiemo Alldieck, Marcus Magnor, Weipeng Xu, Christian Theobalt and Gerard Pons-Moll. Video Based Reconstruction of 3D People Models. In CVPR, 2018.

[14] Mustafa Işık, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, Matthias Nießner. HumanRF: High-Fidelity Neural Radiance Fields for Humans in Motion. In ACM TOG, 2023.

If you see the table below, where 3D reconstruction methods are compared, the performances of methods vary depending on whether the input images are cropped to tightly contain hands or not. PSNR of Livehand using the whole images is 32.60, similar to what they reported. Note in Livehand [2], they use 512 x 334 whole images, where we crop the hand part to reduce the effects of black backgrounds. Prior work [1] also cropped the hand part in their experiments.

Table: Comparison of 3D reconstruction methods using InterHand2.6M. The original high-resolution images from InterHand2.6M, not super-resolved by the proposed method, were used as input.

We hope this addresses your concern.

References:
[1] Xingyu Chen, Baoyuan Wang, and Heung-Yeung Shum. Hand avatar: Free-pose hand animation and rendering from monocular video. In CVPR, 2023.

[2] Ashay Mundra, Mallikarjun B R, Jiayi Wang, Marc Habermann, Christian Theobalt, and Mohamed Elgharib. Livehand: Real-time and photorealistic neural hand rendering. In ICCV, 2023.

As described in the main paper L226 and supplementary L17, we use 20 views x 20 frames for training, and the remaining frames are used for evaluation following prior work [5]. All compared methods (HARP, LiveHand, XHand, UHM) are learned and evaluated on our same setting using the metrics (PSNR, LPIPS, SSIM, P2P).
Note that detailed settings vary in prior works [1, 2, 3, 4, 5], and our number of views is similar to those of LISA, XHand, and HandNeRF, where about 20 cameras are used.

In case the reviewer was inquiring about the evaluation metrics of the methods with the SR module at 512x334 scale images, we report those accuracies in the table below, i.e., HARP, LiveHand, XHand, UHM, and Ours.

References:

[1] Akshay Mundra, Mallikarjun B R, Jiayi Wang, Marc Habermann, Christian Theobalt, and Mohamed Elgharib. Livehand: Real-time and photorealistic neural hand rendering. In ICCV, 2023.

[2] Xingyu Chen, Baoyuan Wang, and Heung-Yeung Shum. Hand avatar: Free-pose hand animation and rendering from monocular video. In CVPR, 2023.

[3] Z. Guo, and et al. HandNeRF: Neural Radiance Fields for Animatable Interacting Hands. In CVPR, 2023.

[4] E. Corona, and et al. LISA: Learning Implicit Shape and Appearance of Hands. In CVPR, 2022.

[5] Qijun Gan, Zijie Zhou, and Jianke Zhu. Xhand: Real-time expressive hand avatar. arXiv:2407.21002, 2024.