Evaluation results on nuScenes

[sanjay-810]

I have been working on evaluating DUST3R on the nuScenes mini dataset, following the details mentioned in your paper (DRIV3R). However, the results I obtained do not align with the reported metrics. Metrics are overshooting and some metrics are infinity.

I would like to confirm if DUST3R needs to be specifically trained on nuScenes (similar to DRIV3R) for evaluation, or if it is expected to generalize without additional training.

For my evaluation:

I used the provided pretrained model of DUST3R. Converted the output of DUST3R into required format for DRIV3R evaluation.
The evaluation protocol was consistent with what was detailed in your github code for the nuScenes dataset.
I would greatly appreciate any clarification or guidance on this matter. If there are additional instructions or training requirements, kindly let me know. And it would be great if you share evaluation script for DUST3R on nuscenes and confirm if DUST3R needs to be specifically trained on nuScenes.

[Barrybarry-Smith]

Though we fine-tuned the DUSt3R to get the metrics in our paper, the results on the original DUSt3R model should not align with what you provided. I strongly recommend to visualize the point cloud first of all models to verify whether all points are in a consistent coordinate system and within the proper range.