Jiezhang Cao
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Jiezhang Cao has a total of 79 co-authors. Here are the top co-authors by number of shared publications:
- Kai Zhang
Nanjing University, Nanjing, China
6 shared publications - Mingkui Tan
South China University of Technology, Guangzhou, China
5 shared publications - Farong Wen
Shanghai Jiao Tong University, Shanghai, China
4 shared publications - Jingyun Liang
Computer Vision Lab ETH Zürich Switzerland
4 shared publications - Luc Van Gool
INSAIT, Sofia University “St. Kliment Ohridski
4 shared publications - Wenguan Wang
Zhejiang University
4 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 16075-16084
Who is a Better Puppeteer: Subjective and Objective Quality Assessment for Generative 3D Animated Human
Although recent advances in AI-based Text-to-Motion (T2M) algorithms have enabled text-driven 3D Animated Human (3DAH), often referred to as "Puppeteers" for their ability to streamline 3D animation design, the quality of the generated motion remains a critical concern. More importantly, the T2M field continues to lack reliable and objective evaluation metrics, which substantially hinders the advancement of puppeteers and negatively impacts user experience. To address this gap, we introduce the first large-scale quality assessment dataset for generative 3DAHs, termed AHQA-3D, which comprises 1,200 3DAHs sequences generated by 6 representative puppeteers. Through carefully designed subjective experiments involving recruited participants, we obtain human quality ratings for each 3DAH. The results highlight the impact of different factors on animation quality. We further propose ViSCA, an objective quality assessment agent based on Qwen2.5-VL. ViSCA supports end-to-end quality evaluation across four dimensions: 3D motion fidelity, 2D human body representation, visual quality, and semantic consistency. Experimental results demonstrate that ViSCA outperforms existing evaluation methods and exhibits stronger alignment with human perception.
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 16075-16084
Open AccessWho is a Better Puppeteer: Subjective and Objective Quality Assessment for Generative 3D Animated Human
Although recent advances in AI-based Text-to-Motion (T2M) algorithms have enabled text-driven 3D Animated Human (3DAH), often referred to as "Puppeteers" for their ability to streamline 3D animation design, the quality of the generated motion remains a critical concern. More importantly, the T2M field continues to lack reliable and objective evaluation metrics, which substantially hinders the advancement of puppeteers and negatively impacts user experience. To address this gap, we introduce the first large-scale quality assessment dataset for generative 3DAHs, termed AHQA-3D, which comprises 1,200 3DAHs sequences generated by 6 representative puppeteers. Through carefully designed subjective experiments involving recruited participants, we obtain human quality ratings for each 3DAH. The results highlight the impact of different factors on animation quality. We further propose ViSCA, an objective quality assessment agent based on Qwen2.5-VL. ViSCA supports end-to-end quality evaluation across four dimensions: 3D motion fidelity, 2D human body representation, visual quality, and semantic consistency. Experimental results demonstrate that ViSCA outperforms existing evaluation methods and exhibits stronger alignment with human perception.
- Open Access
- 4 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
Subjective and Objective QoE Assessment for Real-Time Streaming Interactive Digital Human
Web3 technologies enable novel forms of real-time digital human streaming media by supporting both high-fidelity transmission and interactive user engagement. However, Real-Time Streaming Interactive Digital Humans (RTSIDHs) remain vulnerable to network instability, resulting in buffering, latency, visual degradation, and audio–video desynchronization that substantially impair user Quality of Experience (QoE). To effectively perceive these distortions, we present RDHQA, the first large-scale RTSIDH Quality Assessment dataset. RDHQA comprises 134 representative interaction scenarios with eight digital human avatars as high-quality references, along with 1,340 distorted samples generated by simulating five common streaming degradations. Based on extensive subjective evaluations, we further propose SAV-PF, an audio–visual quality assessment method built on the human foundation model Sapiens and informed by cognitive principles such as the primacy effect and forgetting curve. Experimental results demonstrate that SAV-PF achieves superior performance over existing objective QoE assessment approaches, providing a more accurate prediction of user experience. This work is open-sourced at https://github.com/zyj-2000/RDHQA under the CC BY-NC 4.0 licence.
- 4 June 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Multimedia Computing, Communications, and Applications
Open AccessSubjective and Objective QoE Assessment for Real-Time Streaming Interactive Digital Human
Web3 technologies enable novel forms of real-time digital human streaming media by supporting both high-fidelity transmission and interactive user engagement. However, Real-Time Streaming Interactive Digital Humans (RTSIDHs) remain vulnerable to network instability, resulting in buffering, latency, visual degradation, and audio–video desynchronization that substantially impair user Quality of Experience (QoE). To effectively perceive these distortions, we present RDHQA, the first large-scale RTSIDH Quality Assessment dataset. RDHQA comprises 134 representative interaction scenarios with eight digital human avatars as high-quality references, along with 1,340 distorted samples generated by simulating five common streaming degradations. Based on extensive subjective evaluations, we further propose SAV-PF, an audio–visual quality assessment method built on the human foundation model Sapiens and informed by cognitive principles such as the primacy effect and forgetting curve. Experimental results demonstrate that SAV-PF achieves superior performance over existing objective QoE assessment approaches, providing a more accurate prediction of user experience. This work is open-sourced at https://github.com/zyj-2000/RDHQA under the CC BY-NC 4.0 licence.
- 17 April 2026
- Published by Elsevier in Information Processing & Management
- Vol. 63 (7)
An All-in-One Quality Assessment Agent for 4D digital human: Bridging talking heads and animated human
- 17 April 2026
- Published by Elsevier in Information Processing & Management
- Vol. 63 (7)
An All-in-One Quality Assessment Agent for 4D digital human: Bridging talking heads and animated human
- 20 January 2026
- Published by Elsevier in Pattern Recognition
- Vol. 175
Amplitude exchanging network for unsupervised underwater image enhancement
- 20 January 2026
- Published by Elsevier in Pattern Recognition
- Vol. 175
Amplitude exchanging network for unsupervised underwater image enhancement
- 20 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-12
Degradation-Noise-Aware Deep Unfolding Transformer for Hyperspectral Image Denoising
Hyperspectral Images (HSIs) play a pivotal role in fields like medical diagnosis and agriculture. However, it often contends with significant noise stemming from narrow-band spectral filtering. Existing denoising techniques have their limitations: model-driven methods rely on manual priors and hyperparameters, while learning-based methods struggle to discern intrinsic noise patterns, as they require paired images with specific example noise for training, fail to capture critical noise distribution information, leading to unrobust denoising results. This work addresses the issue by presenting a Degradation-Noise-Aware Unfolding Network (DNA-Net). Unlike training directly with simulated noise, DNA-Net initially models general sparse and Gaussian noise through statistic distributions. It then explicitly represents image priors with a customized spectral transformer. The model is subsequently unfolded into an end-to-end network, with hyperparameters adaptively estimated from noisy HSI and degradation models, effectively regulating each iteration. Furthermore, a novel U-Shaped Local-Non-local-Spectral Transformer (U-LNSA) is introduced, simultaneously capturing spectral correlations, local features, and non-local dependencies. The integration of U-LNSA into DNA-Net establishes the first Transformer-based deep unfolding method for HSI denoising. Experimental results on synthetic and real noise validate DNA-Net’s superior performance over state-of-the-art (SOTA) methods. Moreover, the DNA-Net, trained exclusively on mixed Gaussian noise and impulse noise, demonstrates the ability to generalize to unseen noise present in real images. Code and models will be released at https://github.com/NavyZeng/DNA-Net.
- 20 February 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-12
Degradation-Noise-Aware Deep Unfolding Transformer for Hyperspectral Image Denoising
Hyperspectral Images (HSIs) play a pivotal role in fields like medical diagnosis and agriculture. However, it often contends with significant noise stemming from narrow-band spectral filtering. Existing denoising techniques have their limitations: model-driven methods rely on manual priors and hyperparameters, while learning-based methods struggle to discern intrinsic noise patterns, as they require paired images with specific example noise for training, fail to capture critical noise distribution information, leading to unrobust denoising results. This work addresses the issue by presenting a Degradation-Noise-Aware Unfolding Network (DNA-Net). Unlike training directly with simulated noise, DNA-Net initially models general sparse and Gaussian noise through statistic distributions. It then explicitly represents image priors with a customized spectral transformer. The model is subsequently unfolded into an end-to-end network, with hyperparameters adaptively estimated from noisy HSI and degradation models, effectively regulating each iteration. Furthermore, a novel U-Shaped Local-Non-local-Spectral Transformer (U-LNSA) is introduced, simultaneously capturing spectral correlations, local features, and non-local dependencies. The integration of U-LNSA into DNA-Net establishes the first Transformer-based deep unfolding method for HSI denoising. Experimental results on synthetic and real noise validate DNA-Net’s superior performance over state-of-the-art (SOTA) methods. Moreover, the DNA-Net, trained exclusively on mixed Gaussian noise and impulse noise, demonstrates the ability to generalize to unseen noise present in real images. Code and models will be released at https://github.com/NavyZeng/DNA-Net.
- 14 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 46 (12) , 8365-8379
Towards Lightweight Super-Resolution With Dual Regression Learning
Deep neural networks have exhibited remarkable performance in image super-resolution (SR) tasks by learning a mapping from low-resolution (LR) images to high-resolution (HR) images. However, the SR problem is typically an ill-posed problem and existing methods would come with several limitations. First, the possible mapping space of SR can be extremely large since there may exist many different HR images that can be super-resolved from the same LR image. As a result, it is hard to directly learn a promising SR mapping from such a large space. Second, it is often inevitable to develop very large models with extremely high computational cost to yield promising SR performance. In practice, one can use model compression techniques to obtain compact models by reducing model redundancy. Nevertheless, it is hard for existing model compression methods to accurately identify the redundant components due to the extremely large SR mapping space. To alleviate the first challenge, we propose a dual regression learning scheme to reduce the space of possible SR mappings. Specifically, in addition to the mapping from LR to HR images, we learn an additional dual regression mapping to estimate the downsampling kernel and reconstruct LR images. In this way, the dual mapping acts as a constraint to reduce the space of possible mappings. To address the second challenge, we propose a dual regression compression (DRC) method to reduce model redundancy in both layer-level and channel-level based on channel pruning. Specifically, we first develop a channel number search method that minimizes the dual regression loss to determine the redundancy of each layer. Given the searched channel numbers, we further exploit the dual regression manner to evaluate the importance of channels and prune the redundant ones. Extensive experiments show the effectiveness of our method in obtaining accurate and efficient SR models.
- 14 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Pattern Analysis and Machine Intelligence
- Vol. 46 (12) , 8365-8379
Towards Lightweight Super-Resolution With Dual Regression Learning
Deep neural networks have exhibited remarkable performance in image super-resolution (SR) tasks by learning a mapping from low-resolution (LR) images to high-resolution (HR) images. However, the SR problem is typically an ill-posed problem and existing methods would come with several limitations. First, the possible mapping space of SR can be extremely large since there may exist many different HR images that can be super-resolved from the same LR image. As a result, it is hard to directly learn a promising SR mapping from such a large space. Second, it is often inevitable to develop very large models with extremely high computational cost to yield promising SR performance. In practice, one can use model compression techniques to obtain compact models by reducing model redundancy. Nevertheless, it is hard for existing model compression methods to accurately identify the redundant components due to the extremely large SR mapping space. To alleviate the first challenge, we propose a dual regression learning scheme to reduce the space of possible SR mappings. Specifically, in addition to the mapping from LR to HR images, we learn an additional dual regression mapping to estimate the downsampling kernel and reconstruct LR images. In this way, the dual mapping acts as a constraint to reduce the space of possible mappings. To address the second challenge, we propose a dual regression compression (DRC) method to reduce model redundancy in both layer-level and channel-level based on channel pruning. Specifically, we first develop a channel number search method that minimizes the dual regression loss to determine the redundancy of each layer. Given the searched channel numbers, we further exploit the dual regression manner to evaluate the importance of channels and prune the redundant ones. Extensive experiments show the effectiveness of our method in obtaining accurate and efficient SR models.
- 29 May 2024
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 132 (11) , 4992-5013
Inheriting Bayer’s Legacy: Joint Remosaicing and Denoising for Quad Bayer Image Sensor
- 29 May 2024
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 132 (11) , 4992-5013
Inheriting Bayer’s Legacy: Joint Remosaicing and Denoising for Quad Bayer Image Sensor
- Open Access
Towards Lightweight Super-Resolution with Dual Regression Learning
Deep neural networks have exhibited remarkable performance in image super-resolution (SR) tasks by learning a mapping from low-resolution (LR) images to high-resolution (HR) images. However, the SR problem is typically an ill-posed problem and existing methods would come with several limitations. First, the possible mapping space of SR can be extremely large since there may exist many different HR images that can be super-resolved from the same LR image. As a result, it is hard to directly learn a promising SR mapping from such a large space. Second, it is often inevitable to develop very large models with extremely high computational cost to yield promising SR performance. In practice, one can use model compression techniques to obtain compact models by reducing model redundancy. Nevertheless, it is hard for existing model compression methods to accurately identify the redundant components due to the extremely large SR mapping space. To alleviate the first challenge, we propose a dual regression learning scheme to reduce the space of possible SR mappings. Specifically, in addition to the mapping from LR to HR images, we learn an additional dual regression mapping to estimate the downsampling kernel and reconstruct LR images. In this way, the dual mapping acts as a constraint to reduce the space of possible mappings. To address the second challenge, we propose a dual regression compression (DRC) method to reduce model redundancy in both layer-level and channel-level based on channel pruning. Specifically, we first develop a channel number search method that minimizes the dual regression loss to determine the redundancy of each layer. Given the searched channel numbers, we further exploit the dual regression manner to evaluate the importance of channels and prune the redundant ones. Extensive experiments show the effectiveness of our method in obtaining accurate and efficient SR models.
Open AccessTowards Lightweight Super-Resolution with Dual Regression Learning
Deep neural networks have exhibited remarkable performance in image super-resolution (SR) tasks by learning a mapping from low-resolution (LR) images to high-resolution (HR) images. However, the SR problem is typically an ill-posed problem and existing methods would come with several limitations. First, the possible mapping space of SR can be extremely large since there may exist many different HR images that can be super-resolved from the same LR image. As a result, it is hard to directly learn a promising SR mapping from such a large space. Second, it is often inevitable to develop very large models with extremely high computational cost to yield promising SR performance. In practice, one can use model compression techniques to obtain compact models by reducing model redundancy. Nevertheless, it is hard for existing model compression methods to accurately identify the redundant components due to the extremely large SR mapping space. To alleviate the first challenge, we propose a dual regression learning scheme to reduce the space of possible SR mappings. Specifically, in addition to the mapping from LR to HR images, we learn an additional dual regression mapping to estimate the downsampling kernel and reconstruct LR images. In this way, the dual mapping acts as a constraint to reduce the space of possible mappings. To address the second challenge, we propose a dual regression compression (DRC) method to reduce model redundancy in both layer-level and channel-level based on channel pruning. Specifically, we first develop a channel number search method that minimizes the dual regression loss to determine the redundancy of each layer. Given the searched channel numbers, we further exploit the dual regression manner to evaluate the importance of channels and prune the redundant ones. Extensive experiments show the effectiveness of our method in obtaining accurate and efficient SR models.