We are proud to announce that four groundbreaking research papers authored by Dr. Ling Shao of Terminus Group and his collaborators have been accepted by ICCV 2025. These works advance the frontiers of 3D spatial intelligence, multimodal learning, and scalable medical image analysis, reinforcing our leadership in AI-driven visual computing.
Key Contributions
1. Towards Scalable Spatial Intelligence via 2D-to-3D Data Lifting
This work addresses a critical bottleneck in AI, the scarcity of 3D training data, by introducing a scalable pipeline that lifts single-view images into realistic 3D representations. This approach significantly reduces the cost of 3D data acquisition and enables more widespread development of AI systems capable of understanding and interacting with physical environments. By releasing COCO-3D and Objects365-v2-3D (the first large-scale generated 3D datasets), the method demonstrates versatile applications, from basic perception to multimodal reasoning, amplifying its impact on both academic and industrial spatial tasks.
2. PCR-GS: COLMAP-Free 3D Gaussian Splatting via Pose Co-Regularizations
To eliminate reliance on COLMAP, a longstanding dependency in 3D scene reconstruction, the team introduces PCR-GS, an innovative COLMAP-free 3DGS technique. By aligning semantic features across views and optimizing camera poses via wavelet-frequency analysis, PCR-GS achieves robust 3D modeling even under extreme camera movements. The method demonstrates superior performance in real-world scenarios, marking a practical advancement in the scalability of pose-free 3D generation.
3. Spatial Preference Rewarding for MLLMs’ Spatial Understanding
This paper contributes to improving fine-grained spatial reasoning in multimodal large language models (MLLMs), which often struggle with detailed region description and accurate object localization. The authors propose a novel Spatial Preference Rewarding (SPR) framework that scores and refines model-generated text based on semantic quality and localization accuracy. By comparing low- and high-quality outputs, SPR enables preference-based optimization that enhances spatial alignment with minimal training overhead. The approach demonstrates consistent performance gains across referring and grounding benchmarks, making it an effective method to boost spatial awareness in MLLMs.
4. Scaling Tumor Segmentation: Best Lessons from Real and Synthetic Data
This paper makes a major contribution to tumor segmentation by showing that synthetic data can reduce the need for large amounts of real annotated scans. The team found that syn-thetic augmentation steepens the scaling laws, enhancing Al perormance more efficiently than real data alone. Based on this insight, the authors release CancerVerse, the largest per-voxel annotated tumor dataset across six organs. Models trained on CancerVerse show strong generalization, with up to +16% improvement in out-of-distribution segmentation and +7% improvement in in-distribution segmentation, setting a new benchmark for scalable medical AI.
Dr. Ling Shao, Chief Scientist of Terminus Group, stated: “These acceptances reflect our team’s commitment to solving real-world challenges through fundamental innovations in computer vision. From spatial intelligence to healthcare AI, we’re pushing the boundaries of what’s possible with visual computing.”