acceptodds
Withdrawn from ICLR 2027

GeoX-2: High-Throughput Driving Geometry through Geometry-Preserving Compression

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Feed-forward visual geometry models jointly recover camera pose, depth, and dense 3D structure, but their computational cost limits efficient scaling to synchronized driving views. We present , which compresses the complete inference graph for efficient driving geometry. Structurally, an Efficient Hybrid Encoder reduces the token count of each view, while a lightweight dense head supports efficient geometry inference. To compress the decoder, GeoX-2 treats each Frame–Global block pair as a removable unit and applies Geometry-Derived Structural Slimming. The Pair Importance Score combines LayerScale-based structural saliency with depth-intervention sensitivity to retain important pairs from the full decoder. Prior-free geometry distillation then transfers geometry from a prior-conditioned teacher to the selected compact RGB-only student, removing the conditional adapters and learned point head while recovering 3D points from predicted depth and camera poses. Extensive experiments across multiple driving benchmarks demonstrate strong geometry accuracy. GeoX-2 reaches 1003.1 FPS offline sequence-batched throughput on KITTI and provides leading throughput and peak-memory efficiency across multi-view inference settings with varying view counts.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
VGGT, Driving Scene Reconstruction, Autonomous Driving

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_8CfEOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents GeoX-2, a feed-forward model for efficient driving-scene geometry estimation. It combines a hybrid encoder, decoder-pair pruning, and geometry distillation from a prior-conditioned teacher to an RGB-only student. The model achieves competitive geometry accuracy and 1003.1 FPS on KITTI.

Critical strengths
  1. The paper tackles an important practical problem: improving geometry inference efficiency.
  2. GeoX-2 achieves a good balance between accuracy, speed, and memory usage.
  3. The ablations are useful, particularly the distillation experiment, which reduces Waymo depth Abs Rel from 0.193 to 0.110.
Critical weaknesses
  1. The gain from pruning is unclear. In Table 5, most of the speedup comes from replacing the encoder. Without accuracy results for intermediate models, it is difficult to judge how much pruning actually contributes.
  2. PIS is tested under limited settings. Table 4 only evaluates one pruning budget on Waymo. It is unclear whether the method works equally well when more or fewer decoder pairs are retained.
  3. The role of geometric priors in distillation is not established. The paper shows that teacher supervision helps, but does not compare a prior-conditioned teacher with an RGB-only teacher. This leaves the main source of improvement unclear.
  4. The camera-pose results may be influenced by familiar camera rigs. Since the evaluation datasets also appear in the training list, the model may benefit from known camera configurations. A mean-relative-calibration baseline would help address this concern.
  5. More qualitative results are needed. The paper lacks direct visual comparisons with baselines and ground truth, making it hard to assess local geometric errors and the claimed reduction in checkerboard artifacts. Some figures are also difficult to read due to small labels and unclear connections.
Questions for authors

GeoX-2 is a promising system, especially in terms of efficiency. However, the paper does not clearly establish the individual contributions of pruning and distillation. The limited PIS evaluation, potential bias in camera-pose estimation, and lack of qualitative comparisons also leave important questions unanswered. Thus, a weak rejection is proposed.

Reviewer_Ayu2OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents GeoX-2 for efficient visual geometry reconstruction in driving scenes. It combines a hybrid encoder, decoder pruning, geometry distillation, and a lightweight prediction head to improve inference speed while preserving reconstruction quality.

Critical strengths

The goal of accelerating driving geometry models is reasonable.

Critical weaknesses
  1. Unclear reliability of the depth GT. The paper does not clearly explain how the dense depth GT is constructed, whether it includes pseudo-labels, or how label quality is verified. Since these labels directly determine the geometry evaluation, their reliability is essential for assessing the reported results. The authors should describe the construction process and provide evidence of label quality.

  2. Unfair baseline comparisons. The baselines use official checkpoints directly without matched training data and supervision, making it difficult to attribute the reported gains to the proposed method. The authors should train the baselines with the same data and supervision and report fair comparison results. In addition, DVGT’s nuScenes results are substantially worse than those reported in its original paper. The authors should explain where this gap comes from (e.g., GT quality).

  3. Limited technical novelty. The method mainly combines an encoder adapted from LiAuto-MindViT, simple decoder-layer scoring and removal, distillation that reuses GeoX’s relation loss, and a dense head adapted from VGGT-Ω. These are largely straightforward adaptations of existing methods, with limited fundamental novelty.

  4. Reliance on GT alignment. The geometry results use GT alignment, but GT is unavailable in real-world driving. The authors should report results without GT alignment, as DVGT does, to assess practical performance and allow a fair comparison.

  5. Missing geometry results on OpenScene. OpenScene is included in training and underlies the downstream planning evaluation. So the geometry results on OpenScene are important. However, the paper provides only qualitative geometry results on this dataset. Quantitative results and baseline comparisons should be included.

  6. Misplaced emphasis on throughput. For online driving, the key efficiency metric is how quickly the model processes a newly observed synchronized multi-camera frame. Throughput measured by jointly processing many frames does not establish this response time. Although the paper reports multi-view latency, latency should be the main efficiency comparison, evaluated together with geometry accuracy under the same input setting.

Questions for authors

Please address the weaknesses above, particularly the GT quality and the fair comparisons.

Reviewer_zgqjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

GeoX-2 accelerates multi-view driving geometry prediction by replacing GeoX’s image encoder with a lower-token hybrid encoder, selecting six of twelve Frame–Global decoder pairs using structural and depth-response scores, and distilling a prior-conditioned teacher into an RGB-only student. The student predicts depth and camera parameters and recovers 3D points by unprojection. The paper reports strong reconstruction and depth results across driving datasets, together with improved inference throughput.

Critical strengths
  1. The reported efficiency and geometry results are practically meaningful. GeoX-2 improves reconstruction accuracy and completeness over GeoX on all five evaluated driving datasets, while increasing offline KITTI throughput from 223.8 to 1003.1 FPS.
  2. The paper includes controlled comparisons for pair selection and student training. In particular, Table 6 provides evidence that teacher supervision improves the compact student beyond teacher initialization alone.
Critical weaknesses
  1. The main efficiency gain lacks a sufficiently strong token-reduction baseline. In Table 5, replacing the encoder alone raises KITTI throughput from 223.8 to 774.3 FPS, accounting for most of the final speedup. However, the paper does not compare the resulting quality–speed trade-off with simpler alternatives, such as reducing GeoX’s input resolution, or with other token-reduction approaches at comparable token counts or latency. Therefore, it is unclear whether the chosen hybrid encoder offers a specific advantage over established ways of reducing tokens.
  2. The contribution of the proposed compression steps is difficult to disentangle from the encoder and its pretraining. The hybrid encoder is adopted from prior work and is pretrained on approximately 15 billion image–text pairs. Table 5 reports cumulative speed but not geometry quality after each change. Although pair-selection and distillation ablations are provided, the paper does not establish how much of the final quality–efficiency improvement comes from the encoder, pair slimming, and distillation respectively. Since several components build on existing designs, this missing attribution also limits the evidence for the paper’s methodological novelty.
Questions for authors
  1. At matched inference latency or token count, how does GeoX-2 compare with GeoX using lower-resolution input images? How does it compare with other feasible token-reduction methods?
  2. Can the authors report reconstruction, pose, and depth metrics for every configuration in Table 5, alongside the existing speed measurements?
  3. With the hybrid encoder and its pretraining held fixed, how much additional quality–efficiency benefit comes from the proposed pair-selection rule and distillation procedure over simpler alternatives?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.