DrivingRecon: Large 4D Gaussian Reconstruction Model For Autonomous Driving

Strengths:

1. The PD-Block is a technically sound idea to mitigate spatial redundancy and reallocate Gaussians. 2.The paper reports strong performance on benchmark datasets and supports multiple downstream tasks (reconstruction, editing).
2. The framework is feed-forward and generalizable, avoiding per-scene optimization and enabling real-time rendering.

Weaknesses:

1. For the Inference speed reported in Table 1, is it the average time to run all frames of different scenes (whole scene)? Or it is like, it is the total optimization time for the per-scene optimization methods, but it is the per-frame inference time for generalizable feed-forward methods? So DriveRecon's 0.08s is per-frame inference speed including the regression of all attributes (including 2D/3D flow and semantics) of all Gaussians?
2. I also doubt if it is fair and appropriate to compare the "Inference Speed" this way in Table 1. For the "Per-Scene Optimization" methods, the time is not just for inference, it also includes training. The proposed DriveRecon trained generalization with extremely huge computation (24 x A100 x 24days), so I think it is unfair and inappropriate to compare the "Inference Speed" in this way which includes the optimization time of the "per-scene optimization" methods.
3. Following the comments above, I think the generalizable model is not necessary if per-scene optimization is fast enough, and the generalizable model is too difficult to train. There are already some real-time online GS reconstruction method since last year, such as RTGSLAM (https://github.com/MisEty/RTG-SLAM/tree/main) which achieves real-time GS reconstruction without the need for pretraining. I think a better comparison for Table 1 would be including these real-time GS reconstruction methods, which could potentially achieve much faster inference speed than the "generalizable ones". Plus, some methods such as RTGSLAM also proposes GS management algorithm to efficiently grow and prune GS - If they work comparably good then there is no point of taking huge hardware and time resources to train the model. This is why I think the comparison to these works are very important.
4. Writing and formatting issues: Please check the paper contents carefully. There are mismatches in the explanation of terms around line 98–99, and the order of term definitions is also messy. Additionally, typos such as "fellowing" (lines 181, 189) instead of "following" suggest the writing needs more careful proofreading before submission.

++ The contributions are clearly stated Experiments show that the proposed method outperforms state-of-the-art methods on the nuScenes dataset. The related work section includes most of the recent relevant works in the field of Large Reconstruction Models and Driving Scene Reconstruction.

-- The main issue with the paper is that it lacks some details in the method section: There is no details about the image encoder (size, architecture, etc.). Moreover, the image decoder is not described in detail, which makes it difficult to understand how the image features are transformed into Gaussian points. Moreover, the main contribution of the paper: Prune and Dilate Module is not described in detail: what are the hyperparameters, how they are chosen, etc. In the Static and Dynamic Object Decoupling section, we suppose that the dynamic objects are selected based on their class labels, but this is not clearly stated in the paper.

Strengths

1. The PD-Block introduces a novel mechanism for managing Gaussian points by learning to both prune and dilate them, thereby enabling a more flexible, non-pixel-aligned feedforward gaussian representation.
2. The method outperforms existing feedforward driving reconstruction approaches. This article also uses its image encoder to explore pre-training for end-to-end autonomous driving.
3. A static-dynamic separation strategy for unsupervised learning of autonomous driving scene reconstruction.
4. Extensive experiments, including reconstruction speed, rendering image quality, and rendering depth comparison.

Weaknesses

1. The images in Figure 5 and Figure 6 are of low rendering quality. Additionally, the presentation of novel views feels insufficiently clear. For reference, works like FreeVS, StreetCrafter, and ReconDreamer[1,2,3] utilize specific ego-vehicle lane change distances (in meters) for visualization.
2. Regarding scene editing tasks, is it possible to elevate the scope to include modifications of vehicles within the scene like Street Gaussians[4]? The current article solely showcases the capability of adding novel objects (billboards).
3. In Table 1, why is the reconstruction quality of per-scene optimization methods so low? Is it because the training iterations were limited? For PVG, if it didn't use LiDAR information, did it utilize SfM to extract the initial point cloud for the scene?

Reference

[1]Wang, Qitai, et al. "Freevs: Generative view synthesis on free driving trajectory." arXiv preprint arXiv:2410.18079 (2024).

[2]Ni, Chaojun, et al. "Recondreamer: Crafting world models for driving scene reconstruction via online restoration." Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.

[3]Yan, Yunzhi, et al. "Streetcrafter: Street view synthesis with controllable video diffusion models." Proceedings of the Computer Vision and Pattern Recognition Conference. 2025.

[4]Yan, Yunzhi, et al. "Street gaussians: Modeling dynamic urban scenes with gaussian splatting." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

Strengths

1. Originality of Approach: The PD-Block addresses redundancy in multi-view Gaussian fusion and dynamically allocates computational resources, a novel solution to overlapping artifacts and fixed-point limitations in prior work.

2. Comprehensive Evaluation: The paper compares DrivingRecon against a wide range of baselines (per-scene optimization and feed-forward models) and provides ablation studies for PD-Block and dynamic-static decoupling, validating component effectiveness.

3. Efficiency: DrivingRecon achieves faster inference (0.08s) with fewer parameters (53.37M) than sota methods like STORM (100.60M), enabled by shared encoders and range-view decoders. Weaknesses

4. Inadequate Validation for Fast-Moving or Small Dynamic Objects：while dynamic-static decoupling is proposed, the model’s reconstruction quality for fast-moving objects (e.g., motorcycles, pedestrians) or small entities (e.g., bicycles) is not explicitly tested. The optical flow supervision (Ldy) may struggle with high-velocity motions or occlusions.

5. Adaptive Limitations of PD-Block：The PD-Block relies on manually set thresholds (e.g., τ for redundancy pruning) and empirical parameters (e.g., K centers and Z nearest points), lacking dynamic adaptation to varying scene complexities. For instance, fixed thresholds may incorrectly prune critical Gaussian points in dense traffic or retain redundant points in complex object edges, degrading reconstruction fidelity. The nearest-point aggregation strategy struggles to generalize across scene densities (e.g., urban centers vs. suburban roads), potentially causing feature under-fitting in sparse areas or computational redundancy in dense regions.