AI
Pumpire: Unified Benchmark for Metric Distance Estimation
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 07/10/2026, 20:00:00 UTC
Briefing created: 17/11/2026, 22:18:44 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 09/10/2026, 15:18:24 UTC.
Briefings
Pumpire is a new benchmark designed to evaluate 3D foundation models' ability to estimate physical point-pair distances in reconstructed geometry, addressing limitations of previous methods that separate depth and camera calibration. The project collected a diverse dataset of 100 real-world scenes with 6,400 frames, each annotated with physically measured distances, to enable direct assessment of point-to-point errors across image and video levels. Researchers tested 29 baseline configurations of representative 3D models to provide a comprehensive analysis of scale perception capabilities.
Key points
- Pumpire evaluates 3D foundation models' point-to-point distance estimation by directly measuring physical distances in reconstructed geometry, unlike prior methods that separate depth and camera calibration.
- The project collected 6,400 frames from 100 real-world scenes annotated with physically measured distances to enable direct assessment of point-pair errors across image and video levels.
Context
This benchmark addresses a gap in evaluating 3D models' ability to perceive physical scale, as previous protocols overlooked direct point-to-point distance estimation capabilities. The dataset provides a unified platform for comparing image- and video-level 3D foundation models with or without depth priors.
Limitations
The evaluation relies on a single dataset of 6,400 frames from 100 scenes, limiting generalizability. Results are based on 29 baseline configurations and do not include real-world deployment data. Uncertainty exists regarding the robustness of point-pair distance estimates across different model architectures and hardware setups.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 17/11/2026, 22:18:44 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
IPFS archive pending.