Technical Summary: SpaceVista – All-Scale Visual Spatial Reasoning from mm to km

1. Problem Statement

Current advancements in visual spatial reasoning have largely focused on indoor scenes and specific object scales, often relying on labor-intensive 3D scans and manual annotations. Existing models struggle with diverse real-world applications ranging from millimeter-scale manufacturing to kilometer-scale drone sensing. The field faces two primary challenges:

1. Data Limitations: A heavy reliance on indoor 3D scans and the high cost of curating large-scale, "wild" video datasets with precise physical annotations across diverse scales (mm to km).
2. Model Limitations: The absence of all-scale modeling leads to overfitting on single scenes (typically indoor) and knowledge conflicts when models attempt to reason across vastly different scales (e.g., confusing the size of a screw with a building). Current Multimodal Large Language Models (MLLMs) often lack the ability to generalize spatial understanding from micro-objects to macro-scenes.

2. Methodology

The authors propose a holistic solution comprising a new dataset, a benchmark, and a specialized model architecture.

2.1. Dataset Construction: SpaceVista-1M

To address data scarcity and annotation costs, the authors developed an automated pipeline leveraging specialist models to curate SpaceVista-1M, a dataset containing 1 million spatial QA pairs derived from 38,000 video scenes.

• Scale Coverage: The dataset spans five spatial scales: Tiny Tabletop (mm), Tabletop (cm), Indoor (m), Outdoor (10m), and Drone (100m+).
• Task Diversity: It covers 19 diverse tasks, including general spatial reasoning (distance, counting, relation) and scale-specific tasks (e.g., manipulation planning for tabletops, area estimation for drone views).
• Annotation Strategy: The pipeline integrates camera intrinsics/extrinsics, metric depth estimation (using models like Metric3Dv2 and UniDepthV2), and object tracking (SAM 2 + Grounding DINO) to generate structured 2D/3D annotations.
• Quality Control: While the training set uses automated generation with human verification (achieving ~83% global accuracy), the evaluation benchmark relies on rigorous manual measurement and source verification.

2.2. Benchmark: SpaceVista-Bench

To ensure precise evaluation aligned with the physical world, the authors hand-crafted SpaceVista-Bench.

• Construction: This benchmark involves manually recording and measuring real-world objects (for tiny/tabletop scenes) and retrieving authoritative data (e.g., Wikipedia, official design documents) for landmarks and outdoor scenes.
• Scope: It comprises ~1,600 QA pairs across ~300 unique video scenes, covering all five scales.

2.3. Model Architecture: SpaceVista-7B

The proposed model, SpaceVista-7B, is built upon the Qwen2.5-VL-7B foundation and introduces three key innovations to handle all-scale reasoning:

1. Beyond Semantics (Dense Feature Fusion): To overcome the limitations of semantic-only tokenizers, the model integrates dense features from DINOv3 (a self-supervised vision model known for depth and normal representations). These features are fused with the standard visual encoder via cross-attention, providing geometry-based and depth-based spatial understanding.
2. Scale-Aware Experts (LoRA-like): To mitigate "knowledge conflict" where similar visual patterns imply different physical realities at different scales, the model employs a Mixture of Experts (MoE) approach. Specifically, it uses LoRA-like adapters for different scales, dynamically routed by a Scale Router. This allows the model to maintain independent knowledge for each scale while remaining parameter-efficient.
3. Progressive Reward RL: The training strategy incorporates a Reinforcement Learning (RL) phase using Group Relative Policy Optimization (GRPO). The reward function is designed with "anchors" to enforce a human-like reasoning process:
   • Semantic Anchor: Identifying reference objects.
   • Scale Anchor: Estimating the global scene scale.
   • Answer Anchor: Ensuring the final answer is logically derived.
     This progressive reward structure guides the model to traverse reliable Chain-of-Thought (CoT) paths rather than jumping to conclusions.

3. Key Contributions

1. SpaceVista-1M: The first open-source, all-scale spatial reasoning dataset featuring 1M QA pairs across 5 scales and 19 tasks, supporting both Supervised Fine-Tuning (SFT) with rationales and Reinforcement Learning (RL).
2. SpaceVista-7B: A spatial reasoning model that integrates rich geometric information and employs scale-specific experts with a progressive reward training strategy to alleviate cross-scale conflicts.
3. SpaceVista-Bench: A high-precision, human-annotated benchmark spanning all scales, constructed through manual measurement and authoritative source retrieval to provide a reliable evaluation standard.

4. Experimental Results

The authors evaluated SpaceVista-7B against state-of-the-art closed-source (e.g., GPT-5, Gemini-2.5-pro) and open-source models (e.g., InternVL, Qwen2.5-VL, SpaceR) across five benchmarks.

• Performance: SpaceVista-7B demonstrates competitive performance on existing benchmarks (e.g., VSI-Bench) and achieves substantially superior results on the all-scale SpaceVista-Bench.
• Gains: On SpaceVista-Bench, the model outperforms proprietary and open-source baselines by an average of >3%, with a notable ~6% improvement in comprehensive all-scale reasoning compared to open-source models of similar sizes.
• Ablation Studies:
  • Scale Experts: The introduction of scale-specific LoRA experts significantly improved performance, confirming that separating scale-level knowledge mitigates interference.
  • Progressive Rewards: The anchor-based reward design yielded higher performance than unconstrained reasoning paths.
  • Modality: Incorporating DINOv3 features provided greater gains than using VGGT, highlighting the value of self-supervised dense features for spatial tasks.

5. Significance and Claims

The paper claims that SpaceVista represents a foundational step toward robust all-scale spatial intelligence. By bridging the gap from millimeter-scale manipulation to kilometer-scale remote sensing, the work enables intelligent agents to perceive and reason in unconstrained, real-world environments.

The authors position their work as a catalyst for diverse domains, including:

• Micro-scale: Precision manufacturing and medical surgery.
• Macro-scale: Autonomous driving, drone-based sensing, and cartography.

The paper emphasizes that while current models are limited to indoor or single-scale settings, SpaceVista provides the necessary data, benchmark, and architectural paradigm to advance spatial reasoning across the full spectrum of physical scales. The authors note that while their current coverage is extensive, it does not yet fully address extreme scales (e.g., nanometer semiconductor production or 10km+ satellite cartography), leaving room for future expansion.