Pumpire: Unified Benchmark for Metric Distance Estimation
Pumpire tests whether 3D foundation models can recover real point-to-point distances, not just plausible geometry.
The benchmark uses 100 real-world scenes with physically measured point-pair distances. Each scene has 64 frames, making 6,400 frames in the Pumpire-6k dataset. The authors evaluate both image- and video-level 3D foundation models, with and without depth priors. They report experiments across 29 baseline configurations and argue that prior protocols largely missed physical-scale estimation.
Source: HF Daily Papers' note
The benchmark uses 100 real-world scenes with physically measured point-pair distances. Each scene has 64 frames, making 6,400 frames in the Pumpire-6k dataset. The authors evaluate both image- and video-level 3D foundation models, with and without depth priors. They report experiments across 29 baseline configurations and argue that prior protocols largely missed physical-scale estimation.
Source: HF Daily Papers' note
score 5