MindView
A benchmark for evaluating multi-view spatial intelligence generation. It contains 1,500 samples covering 513 indoor scenes and defines seven geometry- and semantics-driven tasks to test a model's spatial consistency and semantically grounded view generation capability.
Timeline (last 90 days)
The MindView benchmark contains 1500 samples from 513 scenes, covering three indoor datasets.
The MindView benchmark designs seven tasks, divided into two categories: geometry-dominated and semantics-dominated.
The research team evaluated eleven mainstream generative models, and overall performance still shows a significant gap from practical use.
In the paired-context analysis, all eleven generators showed an overall score drop under the extended-context setting.
After fine-tuning on the MindView training set, BAGEL's overall score improved by 24.3 points.
The fine-tuned BAGEL shows gains on all seven tasks.
Learning to generate spatially grounded views may in turn enhance spatial understanding, indicating positive transfer between generation and understanding.
All Facts (7)
The MindView benchmark contains 1500 samples from 513 scenes, covering three indoor datasets.
50%UnverifiedThe MindView benchmark designs seven tasks, divided into two categories: geometry-dominated and semantics-dominated.
50%UnverifiedThe research team evaluated eleven mainstream generative models, and overall performance still shows a significant gap from practical use.
50%UnverifiedIn the paired-context analysis, all eleven generators showed an overall score drop under the extended-context setting.
50%UnverifiedAfter fine-tuning on the MindView training set, BAGEL's overall score improved by 24.3 points.
50%UnverifiedThe fine-tuned BAGEL shows gains on all seven tasks.
50%UnverifiedLearning to generate spatially grounded views may in turn enhance spatial understanding, indicating positive transfer between generation and understanding.
50%