Questions on aligning scene predictions to GT scene

[nstar1125]

Thanks for your amazing work!
I've been trying to replace CUT3R backbone and found out that Human3R's scene prediction aligns well with the GT depth while other metric-depth estimation like Depth-Anything 3 fails to do so. I think this alignment is critical to train SMPL's translation parameter. Is there any idea how you managed to align the scene points to the GT?

Fig. 1. Initial sample from Human3R training (Grey: prediction, Green: GT)

Fig. 2. DA3-metric result (Blue: prediction, Green: GT)

[fanegg]

Hi @nstar1125 Thank you for your interest! May I ask which test dataset used here? It looks like BEDLAM. Since CUT3R's training set includes BEDLAM, you would need to fine-tune DA3 on BEDLAM to achieve better alignment.

[nstar1125]

Thanks for the quick reply! Yes this was a sample from BEDLAM dataset.