SpaceVista
All-Scale Visual Spatial Reasoning from mm to km

With the current surge in spatial reasoning research, researchers have made significant progress in understanding indoor scenes, but still struggle with more diverse applications. This paper aims to advance all-scale spatial reasoning by tackling two key challenges: 1)the heavy reliance on indoor 3D scans and labor-intensive annotations for dataset curation; 2)the absence of all-scale modeling, which often leads to overfitting to single scenes. In this paper, we introduce a holistic solution that integrates a structured spatial reasoning knowledge system, scale-aware modeling, and a progressive training paradigm, as the first attempt to broaden the scope of all-scale spatial intelligence. Using a task-specific, specialist-driven automated pipeline, we curate over 38K video scenes across five spatial scales to create SpaceVista-1M, a dataset comprising one million spatial QAs spanning 19 diverse tasks. While specialist models offer valuable domain knowledge, they are often unreliable evaluators. Therefore, we build an all-scale benchmark with precise annotations by manually recording and retrieving videos. Nevertheless, naive training with SpaceVista-1M often yields suboptimal results due to potential knowledge conflicts. Accordingly, we introduce SpaceVista-7B, a spatial reasoning model that accepts inputs beyond semantics and uses scale as an anchor for scale-aware experts and progressive rewards. Finally, extensive evaluations across five benchmarks, including SpaceVista-Bench, demonstrate competitive performance, showcasing strong generalization across all scales and scenarios.

The limited data and performance constraints in existing models necessitate the creation of a dataset with all-scale spatial context. We propose SpaceVista-1M, a diverse, real-world, all-scale reasoning dataset, as the first to the best of our knowledge. SpaceVista-1M primarily comprises diverse spatial reasoning question–answer pairs, with rich semantic (category, rationale), 2D (mask, box, point), and 3D (depth, camera parameters, point cloud) annotations, obtained either natively or through processing. The construction pipeline in the above figure follows the step-by-step procedure of preparing, transforming, and generating to obtain an all-scale dataset by integrating specialized models.

SpaceVista-7B ingests a question with videos and self-supervised dense features, encodes them, projects features to a shared space, and fuses them in an LLM through learnable interaction. A LoRA-like scale expert with a scale-aware router adapts the model to different spatial scales. Training uses reinforcement learning with stepwise rewards to align reasoning and final answers.