Keep It in Mind: User Centric Continual Spatial Intelligence Reasoning in Egocentric Video Streams¶
🕒 Published (v1): 2026-06-13 08:50 UTC · Source: Arxiv · Venue: ICML · link
Ask a follow-up
Open an assistant pre-loaded with this paper's context.
TL;DR¶
UCS-Bench is a new benchmark of 532 egocentric videos (170+ hours, 8.1K QA pairs) that evaluates whether AI can maintain and update user-centric spatial knowledge — object locations relative to the camera wearer's real-time position — as scenes evolve over long horizons. DirectMe, a training-free framework, incrementally builds a pose-anchored spatial scene graph from streaming observations and uses it to answer queries that require recalling out-of-view objects from the user's current egocentric frame. DirectMe substantially closes the gap over raw MLLM baselines, though a ~37-point gap to human performance (91.7%) remains.
Problem¶
Existing spatial benchmarks and MLLMs evaluate static or episodic spatial understanding from third-person views, failing to address user-centric dynamic spatial reasoning: tracking where objects are relative to the camera wearer's real-time heading and location across long streaming egocentric video. When a user turns around, an object's relative direction (e.g., "behind you") changes even if the object has not moved, and no benchmark or method explicitly evaluates this orientation-aware, long-horizon accumulation.
Method¶
DirectMe is a training-free, streaming QA framework with four stages:
- Multimodal processing (1 fps): DepthAnything3 estimates metric depth and camera extrinsics (ego-pose trajectory
{T_{w←e}(t)}); GroundingDINO + SAM2 produce temporally consistent per-object instance tracks with masks and bounding boxes; a lightweight MLLM assigns per-frame semantic scene tags. - Metric-semantic map: All signals are fused into a global world-coordinate map initialized at t=1 and incrementally updated.
- Evolving scene graph G=(V,E): Nodes are persistent object instances (grounded by 3D world-frame positions) plus semantic place nodes; edges encode object–object and object–place spatial relations in the world frame. At query time t, relations are re-rendered in the current egocentric frame via
T_{e←w}(t) = T_{w←e}(t)^{-1}. - Online QA: A structured query intent is parsed from the question; a minimal subgraph and top-K keyframes are retrieved (causally, from observations up to t); world-frame coordinates are transformed to camera-relative directions and distances; the MLLM generates the final answer from the subgraph + keyframes prompt.
Key Contributions¶
- UCS-Bench: First benchmark targeting user-centric continual spatial intelligence; 8,114 manually annotated QA pairs across four categories (Position & Orientation, Trajectory & Movement, Proximity & Reachability, Category & Quantity), with repeated queries at different timestamps and temporal evidence spans; covers indoor/outdoor, short clips to 44h recordings.
- DirectMe: Training-free, causally streaming scene-graph memory framework that integrates depth, camera pose, and semantic tracking to answer ego-centric spatial queries about out-of-view objects.
- Comprehensive evaluation of 18 MLLMs (4B–38B) across proprietary, open-source, spatial-centric, and streaming categories.
- Root-cause analysis identifying tracking failure (40%), ego-pose relation failure (23%), retrieval failure (20%), and reasoning failure (17%) as dominant error sources.
Results¶
- Human accuracy: 91.7%; Random baseline: 25.7%.
- Best pure MLLM: InternVL3.5-38B 50.2% and InternVL3-38B 53.7% overall; Doubao-seed-1.6V (proprietary) 49.6%.
- DirectMe (w/ Qwen3-VL-8B): 54.1% overall — best reported result, surpassing InternVL3-38B despite using an 8B backbone.
- DirectMe (w/ InternVL3-8B): 52.6% overall.
- DirectMe leads on Position & Orientation (47.6% vs. best baseline ~46.5%), Trajectory & Movement (57.2%), and Proximity & Reachability (59.8%).
- Ablation (Table 5): scene-graph retrieval alone drives the gain (baseline Qwen3-VL 44.6% → graph only 54.6%; combining description + graph 54.1%); MLLM-generated textual description alone yields only marginal improvement (44.3%).
- DirectMe's accuracy degrades more slowly than all baselines as the evidence–question temporal gap increases, particularly beyond 3 minutes.
- Models consistently perform better on object-trajectory than ego-trajectory questions; temporal-state tracking failures account for >55% of all errors.
Limitations¶
- ~37-point gap to human performance remains unresolved.
- DirectMe relies on DepthAnything3 accuracy; errors in depth or pose estimation propagate directly into the scene graph.
- Tracking failure is the single largest failure mode (40% of errors), suggesting the SAM2+GroundingDINO pipeline is a bottleneck in long-horizon streaming settings.
- All proprietary models are capped at 50 input frames (GPT-5 API limit), potentially underrepresenting their true long-video capability.
- The framework is described as training-free, but integrating learning-based spatial modules could further close the human gap — this direction is not explored.
Relevance to Vision-Language Models¶
UCS-Bench directly exposes a failure mode of current MLLMs that is orthogonal to standard VQA: models lack the ability to maintain and update an egocentric spatial reference frame across streaming video, even when they possess strong general visual-language reasoning. The finding that large general-purpose MLLMs (e.g., Qwen3-VL-32B) outperform spatial-centric models on this benchmark challenges the assumption that specialized spatial training transfers to user-centric dynamic settings. DirectMe's training-free scene-graph memory architecture — fusing geometric signals (depth, camera pose) with MLLM reasoning — is a practical blueprint for augmenting VLMs with persistent structured memory, a gap increasingly relevant as VLMs are deployed in wearable and embodied AI contexts.