Arrow Research search

Author name cluster

Sizhe Wang

Possible papers associated with this exact author name in Arrow. This page groups case-insensitive exact name matches and is not a full identity disambiguation profile.

5 papers
2 author rows

Possible papers

5

ICLR Conference 2025 Conference Paper

Bootstrapped Model Predictive Control

  • Yuhang Wang
  • Hanwei Guo
  • Sizhe Wang
  • Long Qian
  • Xuguang Lan

Model Predictive Control (MPC) has been demonstrated to be effective in continuous control tasks. When a world model and a value function are available, planning a sequence of actions ahead of time leads to a better policy. Existing methods typically obtain the value function and the corresponding policy in a model-free manner. However, we find that such an approach struggles with complex tasks, resulting in poor policy learning and inaccurate value estimation. To address this problem, we leverage the strengths of MPC itself. In this work, we introduce Bootstrapped Model Predictive Control (BMPC), a novel algorithm that performs policy learning in a bootstrapped manner. BMPC learns a network policy by imitating an MPC expert, and in turn, uses this policy to guide the MPC process. Combined with model-based TD-learning, our policy learning yields better value estimation and further boosts the efficiency of MPC. We also introduce a lazy reanalyze mechanism, which enables computationally efficient imitation learning. Our method achieves superior performance over prior works on diverse continuous control tasks. In particular, on challenging high-dimensional locomotion tasks, BMPC significantly improves data efficiency while also enhancing asymptotic performance and training stability, with comparable training time and smaller network sizes. Code is available at https://github.com/wertyuilife2/bmpc.

EAAI Journal 2025 Journal Article

Four-dimension efficient pixel-frequency transformer for light field spatial and angular super-resolution

  • Sizhe Wang
  • Hao Sheng
  • Ruixuan Cong
  • Da Yang
  • Rongshan Chen
  • Zhenglong Cui

Super-resolution plays an important role in optical engineering, which can offer high-quality image information for vision applications. The light field camera can capture the intensity and direction information of light, providing abundant scene information. However, it faces the trade-off problem of spatial and angular resolution. Benefiting from global context modeling by self-attention mechanism, transformer-based artificial intelligence methods for the light field (LF) super-resolution (SR) provide superior performance. However, in these methods, the four-dimensional (4D) LF is reshaped into a large one-dimensional (1D) sequence to be processed, which leads to a significant increase in computation and memory and raises difficulties of 4D structure reconstruction for LF spatial and angular SR. This paper presents a 4D efficient pixel-frequency transformer (4DEPFT) for spatial and angular SR, which models 4D correlation in the pixel and frequency domains based on the proposed inner-outer transformer with 4D position encoding. Specifically, two parallel branches separately handle LF images and their corresponding frequency maps in 4DEPFT. In the pixel branch, sub-patch-based inner-outer pixel transformers are proposed to extract 4D pixel features inside and outside local windows, which deal with pixels with small and large disparity shifts separately. In the frequency branch, a multi-band-based inner-outer frequency transformer is proposed to reconstruct 4D frequency features from different frequency bands. Finally, an attention-based pixel and frequency feature fusion module is presented to acquire the high-resolution LF. Experimental results on real-world and simulated public datasets indicate that 4DEPFT outperforms other methods with less computational cost and graphics processing unit (GPU) consumption.

NeurIPS Conference 2025 Conference Paper

MaintainCoder: Maintainable Code Generation Under Dynamic Requirements

  • Zhengren Wang
  • Rui Ling
  • Chufan Wang
  • Yongan Yu
  • Sizhe Wang
  • Zhiyu Li
  • Feiyu Xiong
  • Wentao Zhang

Modern code generation has made significant strides in functional correctness and execution efficiency. However, these systems often overlook a critical dimension in real-world software development: \textit{maintainability}. To handle dynamic requirements with minimal rework, we propose \textbf{MaintainCoder} as a pioneering solution. It integrates the Waterfall model, design patterns, and multi-agent collaboration to systematically enhance cohesion, reduce coupling, achieving clear responsibility boundaries and better maintainability. We also introduce \textbf{MaintainBench}, a benchmark comprising requirement changes and novel dynamic metrics on maintenance efforts. Experiments demonstrate that existing code generation methods struggle to meet maintainability standards when requirements evolve. In contrast, MaintainCoder improves dynamic maintainability metrics by more than 60\% with even higher correctness of initial codes. Furthermore, while static metrics fail to accurately reflect maintainability and even contradict each other, our proposed dynamic metrics exhibit high consistency. Our work not only provides the foundation for maintainable code generation, but also highlights the need for more realistic and comprehensive code generation research. Resources: https: //github. com/IAAR-Shanghai/MaintainCoder.

EAAI Journal 2025 Journal Article

Stereo matching on epipolar plane image for light field depth estimation via oriented structure

  • Rongshan Chen
  • Hao Sheng
  • Ruixuan Cong
  • Da Yang
  • Zhenglong Cui
  • Sizhe Wang
  • Wei Ke

Depth estimation plays a pivotal role in civil engineering such as road surface defect detection, as it serves as a valuable tool, offering high-precision and critical information about scene surface geometry. The Light Field captures both spatial and angular information of a scene, enabling precise depth estimation. The Epipolar Plane Image represents a specific 2-dimensional slice of the light field and is characterized by multiple depth-related lines. Previous epipolar plane image-based methods typically estimate depth maps by extracting the optimal slope for each line; however, they often neglect the visual relationships within this representation, leading to inaccuracies. In this paper, we explore the visual relationship of it and propose a novel visual feature, termed Oriented Structure, which can be utilized to compute scene depth. Similar to previous stereo matching-based methods, we design a new epipolar plane image-based cost volume to extract depth cues from this structure. The cost volume combines the occlusion robustness of epipolar plane image-based methods with the noise robustness of stereo matching-based methods, resulting in smoother depth maps with sharper edges. Building on the framework of existing stereo matching networks, we introduce an epipolar plane image-based stereo matching network for light field depth prediction. Finally, we conduct experiments using both synthetic and real datasets, demonstrating that our network produces higher-quality depth maps compared to previous state-of-the-art methods, ranking first (about 1. 405 mean squared error) on the 4-dimensional light field benchmark. Additionally, we also apply our method to defect detection tasks, providing accurate depth information that leads to improved results.

AAAI Conference 2023 Conference Paper

Take Your Model Further: A General Post-refinement Network for Light Field Disparity Estimation via BadPix Correction

  • Rongshan Chen
  • Hao Sheng
  • Da Yang
  • Sizhe Wang
  • Zhenglong Cui
  • Ruixuan Cong

Most existing light field (LF) disparity estimation algorithms focus on handling occlusion, texture-less or other areas that harm LF structure to improve accuracy, while ignoring other potential modeling ideas. In this paper, we propose a novel idea called Bad Pixel (BadPix) correction for method modeling, then implement a general post-refinement network for LF disparity estimation: Bad-pixel Correction Network (BpCNet). Given an initial disparity map generated by a specific algorithm, we assume that all BadPixs on it are in a small range. Then BpCNet is modeled as a fine-grained search strategy, and a more accurate result can be obtained by evaluating the consistency of LF images in this limited range. Due to the assumption and the consistency between input and output, BpCNet can perform as a general post-refinement network, and can work on almost all existing algorithms iteratively. We demonstrate the feasibility of our theory through extensive experiments, and achieve remarkable performance on the HCI 4D Light Field Benchmark.

v2026.09.13