Request for Depth Backbone Model Weights (Appendix A, Table 12)

[mylwx]

Hi! I was particularly impressed by the description of the Depth Backbone in Appendix A (Figure 7 and Table 12). The experiment showed that the Depth Backbone achieved an mIoU of 42.8% compared to 27.6% for the RGB Backbone, demonstrating its strong capability in encoding depth maps. I noticed that this repository already provides the code, but I couldn’t find the pretrained weights for the Depth Backbone as described in Table 12. Would it be possible for you to share the model file for the Depth Backbone? Having access to this model file would greatly help me in reproducing your experiments and applying it to my research. Thank you very much for your time and consideration!

[yinbow]

We encountered some issues retrieving the previous model from the server. We have retrained a DFormer-S pretrained solely on depth of ImageNet for you, which can achieve better results on NYU. Can you please let us know if it meets your needs. For pure depth input, DFormer-S replaces all three RGB channels with depth.

The pretrained weight can be downloaded at link

[mylwx]

Thank you very much for your prompt response and for providing the pretrained weights for the DFormer-S model trained solely on depth data from ImageNet. I greatly appreciate your effort in retraining the model and sharing the download link. This pretrained weight is also very helpful for my research.

Regarding the Depth Backbone mentioned in your paper, specifically in Appendix A (Figure 7 and Table 12), my understanding is that the Depth Backbone was trained exclusively on ImageNet depth data, using a training codebase similar to ResNet101, with ImageNet depth data as the input. Could you please confirm if this is correct? Additionally, would it be possible for you to provide the model file for this specific Depth Backbone?

Thank you again for your time and support!