Tanks and Temples: Benchmarking Large-Scale Scene Reconstruction

ABSTRACT

We present a benchmark for image-based 3D reconstruction. The benchmark sequences were acquired outside the lab, in realistic conditions.

Ground-truth data was captured using an industrial laser scanner. The benchmark includes both outdoor scenes andindoor environments.

High-resolution video sequences are provided as input, supporting the development of novel pipelines that take advantage of video input to increase reconstruction fidelity.

We report the performance of many image-based 3D reconstruction pipelines on the new benchmark. The results point to exciting challenges and opportunities for future work.

Figure

Figure 1

Ground-truth model for the Panther dataset, one of the datasets in the presented benchmark for large-scale scene reconstruction.

Figure 2

Ground-truth model for the Temple dataset.

This scene has an area of 713 square meters and a height of 21 meters.

The point sets for this and other datasets were meshed to create the renderings shown in the paper. As a result, the renderings may exhibit meshing artifacts that are not present in the ground-truth point sets.

Figure 3

Other models from the intermediate group.

Figure 4

Other models from the advanced group.

Figure 5

SOTA results on a number of benchmark datasets.

(a) A frame from the input video sequence.

(b) Reconstruction produced by the bestperforming pipeline, with distance to the ground-truth model coded by color.

(c) The ground-truth model, with per-point distance to the reconstruction coded by color.

CONCLUSION

We have presented anew benchmark for evaluating image-based reconstruction techniques.

The presented benchmark has a number of characteristics that can support the development of new approaches to 3D reconstruction. Video sequences are provided as input, encouraging new ideas that take advantage of temporally dense sampling to increase reconstruction fidelity.

Complete pipelines are evaluated, aiming to support systems that tackle camera localization and dense reconstruction jointly. Bothoutdoor and indoor scenes are included, with the goal of stimulating the development of robust broad-competence systems. We will set up an evaluation server and online leaderboard that can be used by the community to track progress.