M:3L Lab

ACL 2026

Z3D: Zero-Shot 3D Visual Grounding from Images

A universal pipeline for localizing objects in 3D scenes from multi-view images using zero-shot methods, without geometric supervision or object priors.

Publication record

Status
Accepted at ACL 2026
Authors
Nikita Drozdov, Andrey Lemeshko, Nikita Gavrilov, Anton Konushin, Danila Rukhovich, Maksim Kolodiazhnyi
Venue
ACL 2026
Research area
3D Intelligence
First arXiv submission

Abstract / summary

Z3D is a universal zero-shot 3D visual-grounding pipeline that localizes the object described by a natural-language query. It operates on multi-view images, can optionally incorporate camera poses and depth maps, and avoids task-specific geometric supervision or predefined object priors.

Methodology

  • VLM-assisted image and view selection
  • Prompt-based 2D segmentation with SAM3-Agent
  • Zero-shot 3D proposal generation with MaskClustering
  • 2D-to-3D lifting and multi-view proposal aggregation
  • Optional DUSt3R reconstruction for posed or unposed imagery

Datasets and benchmarks

  • ScanRefer
  • Nr3D

Principal results

  • The paper reports state-of-the-art performance among zero-shot methods on ScanRefer and Nr3D at publication time.
  • In the full ScanRefer validation setting without bounding-box or text supervision, Z3D reports overall [email protected] of 54.2 and [email protected] of 46.0.
  • On Nr3D, Z3D reports overall top-1 accuracy of 54.8 in the paper's main zero-shot comparison.
  • The same pipeline supports point-cloud, posed-RGB, and fully unposed-RGB inputs.

Benchmark evidence

BenchmarkTaskInputMetricResultEvidence
ScanReferZero-shot 3D visual groundingGround-truth point cloud; no bounding-box or text supervisionOverall [email protected] / [email protected]54.2 / 46.0Z3D paper, Table 1
Reported as state of the art among zero-shot approaches in the associated paper.
ScanReferZero-shot 3D visual groundingPosed multi-view RGBOverall [email protected] / [email protected]42.8 / 24.8Z3D paper, Table 1
ScanReferZero-shot 3D visual groundingUnposed multi-view RGBOverall [email protected] / [email protected]31.2 / 12.9Z3D paper, Table 1
Nr3DZero-shot 3D visual groundingDepth-aware settingOverall top-1 accuracy54.8Z3D paper, Table 2
The strongest zero-shot result in the paper's main Nr3D table was 54.8 versus 54.3 for SPAZER.

Values are attributed to the associated paper and the stated input and evaluation protocol.

Technologies and methods

  • PyTorch
  • CLIP
  • SAM3
  • Qwen vision-language models
  • DUSt3R

Related M:3L research

Evidence boundary

Benchmark claims and numerical values are paper-reported results for the named protocol. Potential applications are not evidence of completed deployments.