Uncertainty Quality of VGGT: An Analysis on the DTU Benchmark Dataset

A new analysis of the Visual Geometry Grounded Transformer (VGGT) model, which won the Best Paper Award at CVPR 2025, evaluates the quality of its uncertainty predictions on the DTU benchmark dataset, identifying a confidence threshold that improves the trade-off between reconstruction accuracy and completeness [1][2]. The paper, authored by Markus Hillemann and submitted in June 2026, examines VGGT, a feed-forward neural network that predicts camera poses, depth maps, and dense 3D structure directly from multiple images in a single forward pass without iterative optimization [1]. VGGT can process an arbitrary number of views consistently, a capability that opens possibilities for real-time, scalable photogrammetry [1]. High-quality uncertainty estimates are crucial for fostering trust and enabling robust quality assurance in such applications [1]. The analysis focuses on the DTU benchmark dataset, a standard evaluation set for multi-view stereo [4][5]. The researchers evaluated both the point map branch (VGGT-p) and the depth map branch (VGGT-d) [2][3]. They found that a causally meaningful confidence threshold of 2.0 effectively filters the raw outputs and improves the accuracy–completeness trade-off across scenes [2][3]. Empirical evidence from the DTU evaluation showed that the optimal confidence threshold across scenes ranged between 1.2 and 3.5, with a mean value of 1.9 and a standard deviation of 0.6, supporting the choice of 2.0 as a good general starting point [2][3]. The study revealed that filtering based on the confidence threshold had a much stronger effect on the point map branch than on the depth map branch, which suffered from poor confidence predictions and overconfidence [2][3]. When all points were included without filtering, the results were significantly worse, indicating that data points with the lowest confidences should be excluded from further analysis [2][3]. The point map branch was found to be more accurate than the depth map branch for 3D reconstruction when feed-forward results were evaluated without bundle adjustment [2][3]. VGGT's architecture predicts aleatoric uncertainty for each depth and point map, with the uncertainty maps used in the loss function during training and becoming proportional to the model's confidence in its predictions afterward [4][5]. On the DTU dataset, VGGT substantially outperforms the earlier DUSt3R model, reducing the Overall Chamfer Distance score from 1.741 to 0.382, and achieves results comparable to methods that require known ground-truth cameras at test time [4][5]. The authors attribute these performance gains to VGGT's multi-image training scheme, which teaches the model to reason about multi-view triangulation natively rather than relying on ad hoc alignment procedures [4][5]. The paper concludes that enhancing uncertainty quality holds strong potential for further improving the accuracy of VGGT's 3D reconstructions [1].