E3D-Bench: A Benchmark for End-to-End 3D Geometric Foundation Models
Wenyan Cong, Yiqing Liang, Yancheng Zhang, Ziyi Yang, Yan Wang, Boris Ivanovic, Marco Pavone, Chen Chen, Zhangyang Wang, Zhiwen Fan
TL;DR
This paper introduces E3D-Bench, the first comprehensive benchmark for end-to-end 3D geometric foundation models, evaluating their performance across multiple core 3D tasks and datasets to guide future research.
Contribution
It provides a standardized toolkit and evaluation framework for systematically assessing 3D GFMs, addressing the lack of systematic evaluation in the field.
Findings
16 state-of-the-art GFMs evaluated across tasks
Insights into strengths and limitations of current models
Guidelines for future model scaling and optimization
Abstract
Spatial intelligence, encompassing 3D reconstruction, perception, and reasoning, is fundamental to applications such as robotics, aerial imaging, and extended reality. A key enabler is the real-time, accurate estimation of core 3D attributes (camera parameters, point clouds, depth maps, and 3D point tracks) from unstructured or streaming imagery. Inspired by the success of large foundation models in language and 2D vision, a new class of end-to-end 3D geometric foundation models (GFMs) has emerged, directly predicting dense 3D representations in a single feed-forward pass, eliminating the need for slow or unavailable precomputed camera parameters. Since late 2023, the field has exploded with diverse variants, but systematic evaluation is lacking. In this work, we present the first comprehensive benchmark for 3D GFMs, covering five core tasks: sparse-view depth estimation, video depth…
Peer Reviews
Decision·ICLR 2026 Conference Desk Rejected Submission
(1) This is the first benchmark that systematically evaluates modern 3D GFMs in a unified framework across multiple tasks and data domains. It fills a clear gap in the community. (2) Evaluates 17 models across five different geometric tasks, including both feed-forward ViT-based and diffusion-based models. The scope is wide and well-curated. (3) Standardized evaluation protocols, consistent datasets, unified metrics, and fair hardware settings (A100 for all models). The commitment to releasing c
(1) While the benchmark is comprehensive, the paper does not introduce new model architectures or learning paradigms. Its contribution is mainly infrastructural/empirical rather than methodological. (2) The paper provides extensive quantitative benchmarking but lacks deep qualitative or theoretical analysis of why certain models fail. (3) Although point cloud accuracy/completeness is reported, there is limited evaluation on mesh quality, surface continuity, or structural correctness. (4) The ben
Strength: 1. The paper is well-written and easy to follow. 2. It benchmarks 17 recent 3D geometric foundation models (GFMs), covering both feed-forward transformer and diffusion-based architectures. 3. The benchmark provides useful insights into how end-to-end 3D GFMs perform across multiple tasks, helping the community understand model strengths and weaknesses.
Weaknesses 1. The benchmark reuses existing datasets and introduces no new data. Therefore the main contribution is the findings provided to the community. For the finding: There are confounding factors: 1. When E3D-Bench come to the findings that “no single backbone is universally superior” but since those listed work's training objectives, data scales and additional modules are different. Without controlling for these factors, it is not solid to propose the finding. 2. In 4.3 “stronger 2D fe
Following the seminal DUSt3R model, which introduced dense 3D representation prediction in a single feed-forward pass, numerous geometric foundation models (GFMs) have emerged, each proposing enhancements or variations of the original approach. A comprehensive and fair comparison of these methods across shared tasks, benchmark datasets, evaluation protocols, and metrics represents a highly valuable contribution to the research community. Evaluating the models in the context of novel view synth
The paper includes a broad set of leading GFMs in its analysis, which is highly appreciated. However, it is somewhat unfortunate that MUSt3R (CVPR’25,ArXiv:2503.01661) and MV-DUSt3R+ (CVPR’25,ArXiv:2412.06974), two multi-view extensions of DUSt3R, were not considered in the study. Both have publicly released their code and demonstrated notable improvements over DUSt3R. Additionally, MV-DUSt3R+ introduces support for novel view synthesis (NVS) through lightweight prediction heads that regress 3D
This paper provides a thorough evaluation on multiple geometry foundation models on a series of 3D related tasks, and further summarizes several key findings to help future researches. 1.The included GFMs are rich, including pair-based, multi-view-based, image-sequence-based, and even diffusion models. 2.The evaluation is comprehensive, including metrics such as depth estimation accuracy, pose accuracy, and reconstruction accuracy. 3.The findings are insightful for future researches on GFMs.
1. Training cost and deployment cost are metrics of concern for both researchers and industry. Quantitatively evaluating the relationship between performance and these costs is a potential direction for improvement in the E3D Benchmark. Training cost may include factors such as dataset collection and storage cost, GPU hours required for training, and memory consumption. Deployment cost, on the other hand, may include factors such as FLOPs and memory usage per inference. 2. In the network archite
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
Topics3D Surveying and Cultural Heritage · Geological Modeling and Analysis · 3D Modeling in Geospatial Applications