A new benchmark for 3D reconstruction has found a clear gap between familiar and unfamiliar camera paths. Across five real-world scenes, unseen trajectories were associated with worse results than held-out views from represented trajectories for both tested models and across all three reported image-quality measures.
The testbed was introduced to study reconstruction under more realistic capture and evaluation conditions. It makes trajectory type, pose source and intrinsic source explicit parts of the comparison.
A test built around camera paths
The PIVOT v1 dataset contains five real-world scenes captured with a DJI Mini 4 Pro drone. Each scene includes multiple named and typed trajectories, including paths intended to reflect reconstruction-friendly motion, robot-like movement and extrapolation beyond the captured path.
The testbed retains more than one version of the camera information. It includes measured and optimized pose data, plus physical or offline calibration and COLMAP-optimized intrinsics, so pose components can be selected independently. In ordinary terms, the benchmark can vary both the camera path being tested and the calibration values used to represent it.
Researchers evaluated Nerfacto and Splatfacto using SSIM, PSNR and LPIPS, the three reported image-quality metrics for each trajectory. They also calculated directed pose-space Chamfer distance, which describes pose coverage relative to the training poses rather than measuring novel-view difficulty directly. BM1 used 60,000 iterations, while BM2 and BM3 used 30,000 each.
The gap was visible in the main comparison
The broad result was consistent: across all five scenes, unseen trajectories scored worse than seen held-out views for both models on all three image-quality metrics. The size of the gap varied by scene and model, so the result is not a universal forecast for every reconstruction system.
The Church scene illustrates the difference in PSNR. Nerfacto scored 19.53 dB on seen views and 16.47 dB on unseen views. Splatfacto scored 22.16 dB on seen views and 16.86 dB on unseen views.
The pose-distance analysis pointed in the same general direction without producing a simple rule. Seen views were generally closer to the training pose distribution and had lower LPIPS, while unseen trajectories covered a broader range of pose distances and generally had higher LPIPS. The relationship was not strictly monotonic, so pose distance alone did not capture the full difficulty of rendering a new view.
Results varied with pose and calibration choices
A second benchmark family changed the source of pose components. Among the reported runs, OO, the configuration with optimized translation and rotation, was favored. In Church, MM was about 6.7 to 7.6 dB below OO; in Village Street, MM was 6.93 dB below OO for Nerfacto and 11.56 dB below OO for Splatfacto. The ordering of the mixed configurations was not consistent, so the results do not establish that translation or rotation alone is generally dominant.
Calibration also mattered in the reported comparisons. Fixed OpenCV calibration was associated with lower reconstruction quality than COLMAP-optimized intrinsics in every reported scene and model pair, with PSNR reductions ranging from 1.75 dB to 9.33 dB. The effect varied by scene and model and may partly reflect the quality of the particular calibration used.
What the benchmark cannot yet settle
The calibration comparison has a specific caveat: the physical calibration used in BM3 had approximately four pixels of reprojection error, so part of the gap may reflect the quality of that calibration. Some trajectories may also not fully register in SfM, complicating experiments that require optimized poses.
The benchmark's external reach is limited because PIVOT v1 uses five scenes captured on a single DJI Mini 4 Pro platform. The pattern may not transfer directly to other devices, environments, localization systems or motion platforms.
The findings do not establish a universal trajectory-generalization law, and pose-space Chamfer is not a complete predictor of novel-view difficulty. It is best understood here as a descriptive measure of pose coverage.
A resource for further testing
The paper reports source-code release under the MIT License and dataset release under the CC BY-NC 4.0 license, with repository, dataset and container locations listed.
The manuscript is arXiv:2608.25401v1, dated 26 August 2026. The front matter identifies Mary Raymond as an independent researcher, and no funding source or grant is reported.
Paper data and sources
Original title: PIVOT: A Multi-Trajectory Dataset and Testbed for Pose, Intrinsics, and Novel Viewpoint Evaluation in Real-World 3D Reconstruction
Authors: Mary Raymond
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text