CrossView: Can Vision-Language Models Reason Across Cameras?

Sahil Shah; S P Sharan; Harsh Goel; Manvik Pasula; Adithya Hebbalae; Minkyu Choi; Sandeep P. Chinchali

cs.CVAug 16, 2026

CrossView: Can Vision-Language Models Reason Across Cameras?

Authors: Sahil Shah, S P Sharan, Harsh Goel, Manvik Pasula, Adithya Hebbalae, Minkyu Choi, Sandeep P. Chinchali

Organizations: The University of Texas at Austin, Austin TX, USA · Independent Researcher, USA

Abstract

Video understanding benchmarks have long centered on single-camera settings, where modern multi-modal language models achieve strong performance across image and video tasks. Yet, the real world runs on multi-camera networks: autonomous vehicles, security systems, and robots all gather data across many simultaneous views. We argue that this is not simply "more" of the single-camera problem; it is fundamentally different. Multi-camera reasoning requires handling context that scales with the number of views, resolving occlusions visible from only a subset of cameras, judging which views matter, and integrating evidence across perspectives that may overlap or diverge. Current models struggle with exactly these challenges, yet no benchmark systematically targets them. We introduce CrossView, a multi-camera video question-answering benchmark spanning autonomous driving, security surveillance, egocentric/exocentric video, and robotics. Evaluation of proprietary models, such as GPT-5.2, and open-source models, like Qwen3-VL, reveals consistently low accuracy, with open-source models trailing by a wide margin. Performance scales strongly with a model's ability to jointly process multiple viewpoints, positioning CrossView as a rigorous benchmark for multi-camera video. We open-source our code and dataset at https://utaustin-swarmlab.github.io/CrossView.

Figures & tables

Figure 1: CrossView data example illustrating multi-view reasoning . Multiple synchronized camera streams observe the same scene from different viewpoints (egocentric and exocentric). Answering the question requires integrating information across views (e.g., to identify the event that occurs after the saucepan is placed on the stovetop). CrossView evaluates such cross-view reasoning tasks across diverse domains, including robotics, surveillance, autonomous driving, and ego–exo interaction.

Image & Short-Video QA
Benchmark	#Cam	#Q	Duration	Domains	Cross-View	Cam-ID	Multi-domain	Temp	Evt	Spat	Count	Summ
VQAv2 [ 16 ]	1	1.4M	Image	Natural	✗	✗	✗	✗	✗	✓	✓	✗
GQA [ 25 ]	1	22M	Image	Natural	✗	✗	✗	✗	✗	✓	✗	✗
GazeVQA [ 27 ]	3	25K	$\sim$ 2.5 min	Gaze (ego-exo)	✓	✗	✗	✗	✗	✓	✗	✗
TGIF-QA [ 28 ]	1	165K	$<$ 10 s	GIFs	✗	✗	✗	✓	✗	✗	✓	✗
ActivityNet-QA [ 60 ]	1	58K	$\sim$ 3 min	Activities	✗	✗	✗	✓	✗	✓	✓	✗

Table 1: Comparison of CrossView with existing VQA benchmarks. #Cam : number of simultaneous camera views. #Q : number of questions. Cross-View : whether questions require integrating evidence across camera perspectives. Cam-ID : whether the benchmark includes camera identification/selection questions. Multi-domain : whether the benchmark spans multiple domains. Questions : Temp (temporal), Evt (event ordering), Spat (spatial), Count (counting), Summ (summarization).

Figure 2: Dataset question distribution. Left: Questions are distributed among six categories, each contributing 1,000 questions. Right: Distributions across the four source datasets. MEVA and nuScenes contribute the majority of the questions (2,250 and 1,750 respectively), while Ego-Exo4D and AgiBot provide 1,250 and 750 questions.

Figure 3: Question categories in CrossView . The benchmark includes six types of multi-camera reasoning tasks grouped into three categories: temporal reasoning (temporal queries and event ordering), spatial reasoning (counting and spatial relationships), and holistic understanding (best-camera selection and scene summarization). Each question requires integrating information across multiple synchronized camera views and spans diverse domains including robotics (AgiBot), autonomous driving (nuScenes), surveillance (MEVA), and egocentric–exocentric interaction (Ego-Exo4D).

		nuScenes				AgiBot
Family	Model	Counting	Event Ordering	Spatial	Temporal	Temporal	Event Ordering
Qwen	Qwen2.5-3B	32.5	36.8	30.1	32.2	52.0	50.8
	Qwen2.5-7B	27.5	24.4	45.9	21.6	58.8	54.4
	Qwen3-4B	46.6	46.4	28.9	33.8	66.4	49.2
	Qwen3-8B	43.2	41.2	33.1	36.0	69.6	49.2
InternVL	InternVL2-8B	28.8	36.8	37.3	29.2	42.0	40.4

Table 2: Uniform sampling accuracy (%) on the vehicle-centric and robot-centric subsets of CrossView . Results are reported for nuScenes and AgiBot across counting, event ordering, spatial, and temporal reasoning tasks.

		Ego-Exo4D			MEVA
Family	Model	Temporal	Event Ordering	Best Camera	Counting	Event Ordering	Spatial	Temporal	Best Camera
Qwen	Qwen2.5-3B	50.0	48.0	20.6	22.9	21.9	41.8	32.5	27.3
	Qwen2.5-7B	45.2	49.2	25.8	11.1	83.2	17.6	49.2	34.1
	Qwen3-4B	48.8	50.0	33.2	20.4	34.7	49.0	48.9	34.8
	Qwen3-8B	51.6	46.4	28.8	23.8	36.7	40.4	52.1	38.6
InternVL	InternVL2-8B	28.8	43.6	20.2	18.9	41.4	47.3	52.8	36.6

Table 3: Uniform sampling accuracy (%) on the human-centric subsets of CrossView . Results are shown for Ego-Exo4D and MEVA across temporal reasoning, event ordering, counting, spatial reasoning, and best-camera identification tasks.

		R-1				R-2				R-L
Family	Model	nuScenes	Ego-Exo4D	AgiBot	MEVA	nuScenes	Ego-Exo4D	AgiBot	MEVA	nuScenes	Ego-Exo4D	AgiBot	MEVA
Qwen	Qwen2.5-3B	26.3	28.8	34.4	9.5	5.3	4.3	7.9	0.5	15.0	17.7	21.8	7.6
	Qwen2.5-7B	24.7	18.6	33.2	12.3	4.6	3.7	11.4	0.3	15.2	13.1	24.2	11.5
	Qwen3-4B	17.9	25.9	21.8	6.7	4.9	4.3	6.3	1.0	11.8	15.5	14.3	4.9
	Qwen3-8B	16.4	28.0	29.1	6.5	4.5	3.8	7.5	0.8	10.9	16.8	18.4	5.0
InternVL	InternVL2-8B	16.1	23.0	31.4	5.7	4.0	2.6	6.2	0.6	10.8	15.8	21.2	4.2

Table 4: Summarization performance on CrossView via ROUGE scores (

\times 100

) under uniform frame sampling. R-1 and R-2 denote ROUGE-N overlap of unigrams and bigrams, respectively, and R-L denotes ROUGE-L, based on the longest common subsequence. Frame counts match the evaluation settings used in Tables 2 and 3 .

Figure 4: Qualitative failure case. A driving example from CrossView under uniform sampling , with frames drawn independently from each of the six surround-view cameras. The VLM localizes the roadside construction-fence barrier as b) Front of the vehicle , likely because the front-right and back-right views capture it from a near front-on perspective; the correct answer is c) Right of the vehicle .

Dataset	Input	Counting	Temporal	Event Ordering	Spatial
Ego-Exo4D	Best Exo	–	40.7 ( $-$ 4.5)	47.3 ( $-$ 1.9)	–
Ego-Exo4D	Ego	–	40.4 ( $-$ 4.8)	48.8 ( $-$ 0.4)	–
nuScenes	Front	18.7 ( $-$ 8.8)	21.6 ( $\pm$ 0.0)	32.0 ( $+$ 7.6)	41.5 ( $-$ 4.4)

Table 5: Effect of restricting inputs to a single camera on CrossView for Qwen2.5-7B-Instruct. Compared to multi-camera inputs, single-camera views often reduce performance, showing that many questions require evidence from multiple viewpoints. Deltas are relative to the multi-camera baseline for Qwen2.5-7B from Tables 2 and 3 ( green = single-camera better, red = multi-camera better, gray = no change).

Dataset	Strategy	Counting	Temporal	Event Ordering	Spatial
nuScenes	Uniform	27.5	21.6	24.4	45.9
nuScenes	Stitched	38.0 ( $+$ 10.5)	26.2 ( $+$ 4.6)	39.6 ( $+$ 15.2)	46.3 ( $+$ 0.4)
AgiBot	Uniform	–	58.8	54.4	–
AgiBot	Stitched	–	58.4 ( $-$ 0.4)	48.4 ( $-$ 6.0)	–
Ego-Exo4D	Uniform	–	45.2	49.2	–
Ego-Exo4D	Stitched	–	53.6 ( $+$ 8.4)	48.0 ( $-$ 1.2)	–

Table 6: Comparison of uniform and stitched frame sampling on CrossView for Qwen2.5-7B. Stitched sampling improves performance on several tasks requiring cross-view spatial reasoning ( e.g. , counting and event ordering on nuScenes), but gains are dataset-dependent. Deltas are relative to the uniform baseline from Tables 2 and 3 ( green = stitched better, red = uniform better).

VLM Evaluation

Vision-language model evaluation measures multimodal systems’ perception, reasoning, grounding, robustness, and reliability across controlled and real-world tasks.

See all 957 papers

MVVBench: Benchmarking 4D Reasoning in Vision-Language Models
Sep 25, 2026Hyungjin Chung, Byeongjun Park, Joonseok Lee +3Reasoning for VLMs Video Reasoning
EverEx · Korea University · Yonsei University
MVPBench: A Multi-Video Perception Evaluation Benchmark for Multi-Modal Video Understanding
Mar 24, 2026Purui Bai, Tao Wu, Jiayang Sun +3Video Understanding Video-Language Model Evaluation
MAIS & NLPR, Institute of Automation, Chinese Academy of Sciences, Beijing, China · SIST, ShanghaiTech University, Shanghai, China
SYNCR: A Cross-Video Reasoning Benchmark with Synthetic Grounding
May 8, 2026Sara Ghazanfari, Siddharth Garg, Prashanth Krishnamurthy +1Multimodal Grounding Video Reasoning
New York University

Sep 25, 2026J/K move · Enter open · S save

MVVBench: Benchmarking 4D Reasoning in Vision-Language Models

Hyungjin Chung, Byeongjun Park, Joonseok Lee, Hojun Kim+2

EverEx · Korea University · Yonsei University

Reasoning for VLMs Video Reasoning VLM Evaluation 4D Scene Understanding

Open paper AskPDF arXiv

Multi-view video understanding requires integrating spatial and temporal evidence across multiple, often non-overlapping camera streams: tracking entities as they transition between viewpoints, aligning events across time, and reasoning about latent 4D continuity rather than any single visible frame. We introduce MVVBench, a benchmark for multi-view video reasoning built from real world multi camera datasets. Questions are curated to be monocular-ambiguous along both the view and the temporal axis: each question is unanswerable from any single view in the designated input set, and the majority are further unanswerable from any single moment. Each question becomes uniquely solvable only by jointly reasoning across views and across time. MVVBench spans diverse dynamic scenes and probes six capabilities: implicit/explicit attribute identification, implicit/explicit relative distance, relative camera pose, and compositional counting, with human-authored QA and rigorous verification. Beyond benchmarking, we provide an extensive analysis of when and why current vision language models succeed or fail, characterizing errors due to temporal mis-localization, cross-view identity breaks, and brittle multi-hop reasoning. We then study inference-time elicitation strategies that unlock latent multi-view competence---task-specific chain-of-thought scaffolds and structured cross-view evidence aggregation---yielding substantial gains without retraining. Finally, we present preliminary evidence that reinforcement learning with verifiable rewards can elicit some latent multi-view competence in the base model, pointing to training-time approaches as a promising direction for future work. Together, MVVBench offers a rigorous evaluation of 4D multi-view reasoning and a foundation for future progress toward reliable embodied perception.

Abstract

VLM Evaluation

Explore similar work

MVVBench: Benchmarking 4D Reasoning in Vision-Language Models

MVPBench: A Multi-Video Perception Evaluation Benchmark for Multi-Modal Video Understanding

SYNCR: A Cross-Video Reasoning Benchmark with Synthetic Grounding