SpatialBench: Is Your Spatial Foundation Model an All-Round Player?

While spatial foundation models have demonstrated impressive performance on standard datasets, a critical question remains: are they truly all-round players capable of generalizing robustly across diverse downstream tasks, arbitrary viewpoints, shifting scene domains, varying input densities, and specific hardware constraints? Answering this overarching question requires a holistic assessment, yet current models are mainly evaluated on specific domains for which they were specifically designed or trained. Such evaluations are intrinsically limited by narrow paradigm coverage, limited scene domains, and arbitrary frame sampling, making it fundamentally difficult to assess their true generalization capabilities. To address this gap, we present SpatialBench, a cross-paradigm, domain-diverse benchmark for spatial foundation models with deterministic sampling. SpatialBench features unprecedented scale and rigorous deterministic design, comprising 19 datasets and 546 scenes across 5 diverse spatial domains. It comprehensively evaluates 41 models across 6 paradigms on 5 task suites under 4 different input density settings. Our extensive evaluation reveals that current models are not yet all-round players, and uncovers crucial insights for future advancement. Specifically, we demonstrate that full-context attention maximizes accuracy while bounded-memory strategies unlock long-sequence scalability. Moreover, our empirical evaluations in challenging embodied and egocentric tasks demonstrate that strict domain alignment and high data quality are far more critical to performance than simple dataset scaling. Furthermore, to address the largest data gap identified in our analysis, we go beyond evaluation by introducing a large-scale dataset, DA-Next-5M, and a strong baseline model, DA-Next, pushing the boundaries of spatial representation learning.

📋 한 줄 요약

**[Spatial Foundation Model 평가]** SpatialBench가 cross-paradigm·domain-diverse·deterministic 벤치마크로 19 데이터셋·546 scene·41 모델·6 paradigm·4 input density 통합 평가 — 현재 모델은 all-round 아님 확인, full-context attention·strict domain alignment 중요성 정량 입증, DA-Next-5M·DA-Next 추가 공개.

🎯 핵심 기여도

Spatial foundation model이 standard dataset에서 impressive 성능 보이지만, diverse downstream task·arbitrary viewpoint·shifting scene domain·varying input density·specific hardware constraint에 robustly 일반화하는 "all-round player"인지가 critical 질문임을 지적.
현재 모델이 주로 specifically 설계·학습된 도메인에서만 평가됨 — narrow paradigm coverage·limited scene domain·arbitrary frame sampling으로 fundamentally 한계.
SpatialBench 제시 — cross-paradigm·domain-diverse 벤치마크, deterministic sampling. 19 데이터셋·546 scene 5 spatial 도메인의 unprecedented 규모·rigorous deterministic 설계.
6 paradigm 41 모델을 5 task suite·4 input density 설정 하에 종합 평가. Full-context attention이 정확도 최대화·bounded-memory는 long-sequence scalability unlock, embodied·egocentric task에서 strict domain alignment·high data quality가 단순 데이터 스케일링보다 far more critical. 가장 큰 데이터 격차 해소 위해 DA-Next-5M 데이터셋·DA-Next baseline 모델 도입.

💡 핵심 아이디어

Spatial foundation model의 진정한 일반화 능력은 deterministic·cross-paradigm·domain-diverse 평가로만 측정 가능하며, all-round player를 만들려면 단순 scale 확대보다 strict domain alignment와 high data quality가 critical하다.

🔬 기술적 접근법

**방법론**: SpatialBench — cross-paradigm·domain-diverse 벤치마크 + DA-Next-5M 데이터·DA-Next 모델.
**핵심 기법**: (1) 19 데이터셋·546 scene 5 spatial 도메인 통합, (2) 6 paradigm·41 모델·5 task suite·4 input density 설정, (3) Deterministic sampling으로 arbitrary frame sampling 한계 해소, (4) Full-context attention vs bounded-memory 정량 비교, (5) DA-Next-5M 대규모 데이터셋·DA-Next baseline 모델 추가 공개.

📊 주요 결과

현재 spatial foundation model이 all-round player 아님 확인.
Full-context attention이 정확도 최대화, bounded-memory가 long-sequence scalability 제공.
Embodied·egocentric task에서 strict domain alignment·high data quality가 dataset scaling보다 far more critical.
DA-Next-5M·DA-Next로 가장 큰 데이터 격차 해소, spatial representation learning 경계 push.

💭 의의 및 한계

**의의**: Spatial FM 평가의 cross-paradigm·deterministic 표준 정립, scale > quality 통념 반박(quality·domain alignment 우월), 평가 너머 DA-Next-5M 데이터·baseline 모델 동봉으로 분야 추진. **한계**: 41 모델·5 도메인의 커버리지가 미래 paradigm으로 일반화 한계, deterministic sampling의 평가 다양성 trade-off, hardware constraint 평가의 구체적 metric 부족.

🚀 실용적 활용

Spatial FM 개발의 표준 평가.
자율주행·로봇·AR/VR의 spatial reasoning 모델 선택.
데이터 큐레이션 전략 가이드(scale vs quality).