MBZUAI researchers present a 3D language model for fine-grained grounding
MBZUAI’s Ground3D-LMM links text prompts to 3D point clouds to locate and measure objects and their parts.
Event / data period:
Published: · Updated:
Ground3D-LMM, developed by MBZUAI researchers, uses text prompts with 3D point-cloud data to locate and measure particular objects and object parts in a scene. The university’s 20 October feature says the model was trained and evaluated with a dataset built from ScanNet and ScanNet++, with annotations for objects and their parts.
The dataset contains about 2.5 million question-answer pairs across roughly 2,500 indoor environments and eight tasks. MBZUAI reported that Ground3D-LMM outperformed other systems by nearly 20% on ScanRefer without using 2D images; this is a comparative result on a specified research benchmark. The findings were presented at ECCV 2026. The work does not establish real-world robot capability or reliable physical measurement.
Data
Ground3D-LMM dataset summary
Scope of the dataset assembled by the researchers for model training and evaluation.
| Question-answer pairs | 2500000 | count |
| Indoor environments | 2500 | count |
| Tasks | 8 | count |
Methodology
These figures describe the approximate size and scope of a research dataset; they are not Abu Dhabi or UAE measurements. The reported near-20% ScanRefer comparison is qualified in the story and should not be interpreted as general performance or a field result.
Updated:
Subjects and places
Subject proposals
- Science and technologySuggested
- Preserved editorial category/topic; broad candidate onlymbzuai.ac.ae ↗
Geographic references
Only source-provided geographic references are shown. Opening an external map is optional. Optional story maps use OpenStreetMap. Opening a map sends your IP address and this site’s address to the map provider; location data remains readable as text without loading the map.
- United Arab Emirates · Place role: Coverage · Place type: Country or area · Location precision: RegionalEvidence
- Explicit preserved editorial country/region; no city or national centroid inferredmbzuai.ac.ae ↗