Supplementary Material for ReconDreamer++: Harmonizing Generative and Reconstructive Models for Driving Scene Representation

Waymo. For the selection of Waymo [waymo] scenes, we follow the approach used in ReconDreamer [recondreamer]. These eight scenes are characterized by their rich dynamic foregrounds and large-scale variations. The specific scene names are listed in Tab. 1.

nuScenes. Similarly, we select eight scenes from the nuScenes dataset [nuscenes], each featuring complex traffic environments with diverse and challenging conditions. The specific scenes are as follows: 0037, 0040, 0050, 0062, 0064, 0086, 0087, 0100.

PandaSet. The selection of scenes from the PandaSet dataset [pandaset] follows the approach used in UniSim [unisim] and NeuRAD [Neurad]. The specific scenes are 001, 011, 016, 028, 053, 063, 084, 106, 123, and 158.

EUVS. EUVS [euvs] is a dataset specifically designed for extrapolated urban views, where all results across different lanes are captured from real-world observations. However, since the data for different lanes are not collected simultaneously, the dataset primarily focuses on static backgrounds, requiring the foreground to be ignored during evaluation. Consequently, when calculating metrics such as PSNR, SSIM, and LPIPS, the foreground must first be masked out to ensure accurate assessment of the static scene reconstruction. The four scenes from Level 1 are: vegas_location_1, vegas_location_2, vegas_location_15, vegas_location_22.

Metrics. As mentioned in the main text, we utilize Novel Trajectory Agent Intersection over Union (NTA-IoU) and Novel Trajectory Lane Intersection over Union (NTL-IoU) to assess the quality of the rendered video, both metrics proposed in DriveDreamer4D [drivedreamer4d]. These metrics are specifically designed to evaluate the spatiotemporal coherence of foreground agents and background lanes, respectively.

nuScenes. The visualization results for the nuScenes dataset are presented in Fig. 1, offering a comprehensive comparison of the performance of different methods. The experimental findings demonstrate that ReconDreamer++ achieves results on the original trajectory that are comparable to traditional reconstruction techniques such as Street Gaussians [streetgaussian]. Notably, ReconDreamer++ even outperforms Street Gaussians in certain fine-grained details, showcasing its ability to capture intricate structures within the scene. In contrast, ReconDreamer [recondreamer] exhibits limitations, producing relatively blurry and less detailed rendering results on the original trajectory, which highlights the challenges faced by earlier approaches. When evaluating novel trajectories, the visualizations reveal the superior capabilities of ReconDreamer++. Specifically, the method demonstrates exceptional performance in maintaining high geometric consistency of structured elements, particularly the ground surface, with respect to the ground truth. As highlighted by the orange bounding boxes in Fig. 1, ReconDreamer++ accurately renders left-turn arrows on the ground under novel trajectories, maintaining strong geometric fidelity. In contrast, both ReconDreamer and Street Gaussians struggle to render these ground structures with sufficient accuracy, often leading to distorted or incomplete representations. Such inaccuracies can significantly impact downstream tasks in autonomous driving, where precise understanding of road markings and other structured elements is essential for safe navigation.

PandaSet. The visualization results for PandaSet [pandaset] are shown in Fig. 2, with comparisons that align with the experimental findings on Waymo [waymo] and nuScenes [nuscenes]. ReconDreamer++ demonstrates a comprehensive improvement over ReconDreamer [recondreamer] in rendering quality for both original and novel trajectories. Notably, it not only ensures rendering performance on par with traditional reconstruction methods, such as Street Gaussians [streetgaussian], for original trajectories, but also achieves state-of-the-art (SOTA) results for novel trajectories. In particular, ReconDreamer++ excels in rendering structured elements such as lane markings on the ground, achieving exceptional accuracy and consistency. As highlighted by the orange bounding boxes in Fig. 2, the method accurately reconstructs these critical road features even under challenging novel viewpoints. These results further validate the effectiveness and robustness of ReconDreamer++ across diverse datasets and scenarios.

EUVS. The experimental results for EUVS [euvs] are shown in Fig. 3, demonstrating the superior performance of ReconDreamer++ on both the training set and the test set. Since this dataset consists of data collected from the same scene at different times, the dynamic object regions—including cars, pedestrians, and other moving elements—in the rendered results are masked to facilitate a clearer and more accurate comparison. Specifically, as highlighted by the orange bounding boxes in Fig. 3, ReconDreamer++ maintains an exceptionally high level of consistency with the ground truth in terms of detail rendering.