Three papers have been accepted to ECCV2026
For the ECCV 2026 main conference, 10,473 submissions were assigned for review. The Program Committee recommended 2,883 papers for acceptance, resulting in an acceptance rate of 27.53%.
Enhancing Embodied Reasoning and Grounding by Novel View Synthesis
Authors: Kim Yu-Ji (POSTECH), Dahye Lee (KAIST), Kim Jun-Seong (POSTECH), Nam Hyeon-Woo (POSTECH), GeonU Kim (KAIST), Yongjin Kwon (ETRI), Yu-Chiang Frank Wang (NVIDIA), Jaesung Choe (NVIDIA), Tae-Hyun Oh (KAIST)
While recent RGB-D structured memories provide a foundation for scene and spatial understanding, they are inherently constrained by the fixed viewpoints. These constraints lead to occlusions and restricted fields of view, highlighting the necessity of novel-view reasoning to achieve a more comprehensive scene understanding. We present SplatReasoner, an enhanced embodied reasoning and grounding framework that integrates 3D Gaussian Splatting (3DGS) with Vision-Language Models (VLMs). SplatReasoner first identifies the most relevant pre-captured views for a user query, and subsequently adjusts them into novel viewpoints using 3DGS-based view synthesis. This enables VLMs to better align visual information, facilitating their reasoning process. To facilitate 3D spatial understanding, we utilize language-embedded 3DGS for visual grounding. While this approach has shown promise in open-vocabulary tasks, its efficacy is bounded to simple categorical queries and struggles to interpret complex, compositional language. By interacting with our enhanced reasoning outputs, we successfully perform 3D grounding by handling sophisticated linguistic queries. Experiments show that SplatReasoner outperforms in OpenEQA and ScanRefer benchmarks for embodied reasoning and 3D grounding, respectively, demonstrating the effectiveness of combining VLM-based reasoning with 3DGS for 3D scene understanding.
Reflection-aware Generative Novel View Synthesis
Authors: GeonU Kim (KAIST), Shin Dong-Yeon (KAIST), Tae-Hyun Oh (KAIST)
We propose Ref-GeNVS, a training-free, reflection-aware method for generative novel view synthesis (NVS) in mirror scenes. Existing multi-view diffusion models often fail to recognize mirrors in a scene and cannot exploit reflected content for scene generation. To address this issue without additional training, our key idea is to treat a mirror image as two complementary views. From a single or sparse set of inputs, we estimate the mirror plane and reflect camera poses to form virtual views. Based on this virtual-view setup, we propose Mirror-Gated Attention and Reflection Injection, which enable reflection-consistent NVS by explicitly leveraging reflection relationships within a multi-view diffusion model. Ref-GeNVS inherits the strong generalization capability of the underlying multi-view diffusion backbone while requiring no fine-tuning. On both synthetic and real-world scenes containing mirrors, Ref-GeNVS outperforms recent generative NVS methods by producing reflection-consistent and contextually coherent novel views.
SA-ResGS: Self-Augmented Residual 3D Gaussian Splatting for Next Best View Selection
Authors: Kim Jun-Seong, Tae-Hyun Oh (KAIST), Eduardo Pérez-Pellitero (Huawei Noah’s Ark Lab), Youngkyoon Jang (Huawei Noah’s Ark Lab)
We propose Self-Augmented Residual 3D Gaussian Splatting, a novel framework for stabilizing uncertainty quantification and enhancing uncertainty-aware supervision in Next-Best-View selection for active scene reconstruction. To efficiently estimate scene coverage, SA-ResGS generates geometry-consistent Self-Augmented point clouds (SA-Points) via triangulation between observed training views and rasterized extrapolated views. To address the lack of learning signals in underrepresented regions within sparse, wide-baseline settings, we introduce the first skip-connection-inspired residual learning strategy tailored for 3DGS. This mechanism amplifies gradient flow to weakly contributing, high-uncertainty Gaussians. Our contributions are threefold: (1) a physically grounded, diversified view selection strategy; (2) an uncertainty-aware residual supervision scheme that improves gradient flow and learning stability; and (3) implicitly debiased uncertainty quantification resulting from constrained view selection and residual supervision. Experiments on NeRF Synthetic, Mip-NeRF 360, and challenging extended benchmark from Deep Blending and Tanks and Temples demonstrate that SA-ResGS consistently outperforms state-of-the-art competing methods in both reconstruction quality and view selection robustness.