An arXiv preprint reports a computer-vision system that aims to ease the memory burden of multi-view stereo by using fewer depth hypotheses, while posting a competitive result on DTU. The method, called ZipMVS, reported 0.369 mm accuracy, 0.284 mm completeness and 0.327 mm overall error. The paper says that overall error ranked second among the listed methods.
A smaller set of depth guesses
Multi-view stereo uses a cost volume, a memory-heavy store of depth hypotheses tested for each pixel. ZipMVS makes that sampling adaptive at pixel level, and uses a differentiable scaling network to set the depth range. A GRU-based Depth Speculator generates speculative depth hypotheses, with the method designed to work with fewer samples while retaining reconstruction quality.
The evaluation covered three public datasets and an additional in-house synthetic dataset simulating space-application conditions. Training used the DTU training set with 512 × 640 input images, five views, 24 epochs and a batch size of four, using Adam on one NVIDIA RTX 4090 GPU. DTU testing used four input views at 1200 × 1600 resolution.
Competitive results, with trade-offs
On Tanks and Temples, ZipMVS reported an F-score of 54.46 on the Intermediate subset, with precision of 44.77 and recall of 74.79. On the Advanced subset, its F-score was 33.03, with precision of 27.67 and recall of 42.38.
The memory and runtime figures give the clearest picture of the trade-off in the reported DTU comparison. ZipMVS used 1322 MB and took 0.112 seconds. IterMVS used 886 MB and 0.093 seconds, while PatchmatchNet used 1629 MB and 0.139 seconds. ZipMVS therefore sat between those two methods on both reported measures.
On the in-house synthetic space scenario, the paper reports superior reconstruction quality for ZipMVS and memory use of 824 MB. The supplied analysis does not provide a numerical reconstruction-quality metric for that comparison. In a separate timing comparison, ZipMVS took 0.073 seconds per depth map, compared with 0.078 seconds for PatchmatchNet, whose listed memory use was 1024 MB.
Inside the model
In the DTU module ablation, the configuration without both the Adaptive Range and Central Densely modules recorded 0.383 mm overall error. The figures were 0.352 mm without Adaptive Range, 0.364 mm without Central Densely and 0.327 mm for full ZipMVS, the lowest reported value in these configurations.
Another ablation examined the GRU-based Depth Speculator. The configuration with compensated depths at both levels recorded 0.362 mm overall error; speculative depth at level 2 only recorded 0.353 mm, and speculative depth at level 1 only recorded 0.349 mm. The lowest reported figure was 0.340 mm when speculative depths were used at both levels.
Changing the number of input views produced a different pattern in the DTU test. Overall error was 0.3520 mm with two views, 0.3502 mm with three, 0.3266 mm with four, 0.3396 mm with five and 0.3395 mm with six. The lowest reported error occurred at four views, while memory use and runtime increased across the reported view counts.
The boundaries of the evidence
The results should be read as comparisons within the tested settings. The paper reports no uncertainty intervals or inferential tests for the benchmark comparisons. On DTU, the overall-error ranking was second among the listed methods, while the synthetic quality advantage was qualitative rather than backed by a supplied numerical metric.
The document is an arXiv version 1 preprint dated 28 August 2026. The paper says its code is available on GitHub, and the supplied acknowledgments do not report a funding source.
Paper data and sources
Original title: ZipMVS: Multi-View Stereo with Compressed Cost Volumes
Authors: Guanglin Jin, Hongshan Yu, Javier Civera, Zhaoxin Li
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text