SpatioTemporal Difference Network for Video Depth Super-Resolution (Oral)

Zhengxue Wang, Yuan Wu, Xiang Li, Zhiqiang Yan ✉, Jian Yang ✉

• We propose STDNet, a novel framework for video depth super-resolution. STDNet introduces spatial and temporal difference mechanisms to mitigate long-tailed effects in video depth super-resolution. This design enables precise depth calibration and motion compensation, leading to state-of-the-art performance.

Diffusion-Based Contextual Reconstruction for Point Cloud Segmentation with Limited Annotations

Jiawei Lian, Zhengxue Wang, Wentao Qu, Haobo Jiang, Le Hui, Jian Yang

• We introduce DiCoSeg, the first diffusion-based framework for 3D point cloud semantic segmentation under limited annotation settings. DiCoSeg reconstructs contextual semantics from noisy inputs and dynamically aggregates both local and global spatial structures, thereby achieving stronger contextual awareness and robustness.

Event-Driven Dynamic Scene Depth Completion

Zhiqiang Yan, Jianhao Jiao, Zhengxue Wang, Gim Hee Lee

• We introduce EventDC, the first depth completion framework that tackles the challenges of dynamic scenes by harnessing the unique strengths of event data. To mitigate the adverse effects of fast ego-motion and object motion, EventDC incorporates two event-driven modules. Furthermore, to support research in this area, we construct the first benchmark for event-based depth completion comprising one real-world and two synthetic datasets.

DuCos: Duality Constrained Depth Super-Resolution via Foundation Model

Zhiqiang Yan, Zhengxue Wang, Haoye Dong, Jun Li, Jian Yang, Gim Hee Lee

We introduce DuCos, a novel depth super-resolution framework grounded in Lagrangian duality theory, offering a flexible integration of multiple constraints and reconstruction objectives to enhance accuracy and robustness. Our DuCos is the first to significantly improve generalization across diverse scenarios with foundation models as prompts. Crucially, these prompts are seamlessly embedded into the Lagrangian constraint term, forming a synergistic and principled framework.

DORNet: A Degradation Oriented and Regularized Network for Blind Depth Super-Resolution (Oral)

Zhengxue Wang *, Zhiqiang Yan * ✉, Jinshan Pan, Guangwei Gao, Kai Zhang, Jian Yang ✉

For the first time, we introduce a Degradation Oriented and Regularized Network (DORNet) designed for real-world depth super-resolution, addressing the challenges posed by unconventional and unknown degradations. The core concept involves estimating implicit degradation representations to achieve effective RGB-D fusion. This degradation learning process is self-supervised.

Completion as Enhancement: A Degradation-Aware Selective Image Guided Network for Depth Completion

Zhiqiang Yan, Zhengxue Wang, Kun Wang, Jun Li , ✉ Jian Yang ✉

We propose a novel degradation-aware framework SigNet that transforms depth completion into depth enhancement for the first time. SigNet eliminates the mismatch and ambiguity caused by direct convolution over irregularly sampled sparse data. Meanwhile, it builds a self-supervised degradation bridge between coarse depth and targeted dense depth for effective RGB-D fusion.

Deep Height Decoupling for Precise Vision-based 3D Occupancy Prediction

Yuan Wu *, Zhiqiang Yan * ✉, Zhengxue Wang, Xiang Li, Le Hui, Jian Yang ✉

• For the first time, we introduce the explicit height prior into the vision-based 3D occupancy predition task. Owing to the novel deep height decoupling and sampling stratagy, our model achieves state-of-the-art performance even with minimal input cost.

SGNet: Structure Guided Network via Gradient-Frequency Awareness for Depth Map Super-Resolution

Zhengxue Wang, Zhiqiang Yan ✉, Jian Yang ✉

SGNet introduces a novel perspective that exploits the gradient and frequency domains for the structure enhancement of DSR task, surpassing the five state-of-the-art methods by 16% (RGB-D-D), 24% (Middlebury), 21% (Lu) and 15% (NYU-v2) in average.

Lightweight Bimodal Network for Single-Image Super-Resolution via Symmetric CNN and Recursive Transformer (Short Oral)

Guangwei Gao *, Zhengxue Wang *, Juncheng Li, Wenjie Li, Yi Yu, Tieyong Zeng

LBNet introduces a lightweight bimodal network that integrates a CNN for local feature extraction with a recursive Transformer for global dependency modeling. It reduces computational and memory costs while more effectively enhancing texture details in single-image super-resolution.