search

cnexah / VA-DepthNet

VA-DepthNet: A Variational Approach to Single Image Depth Prediction
MIT License
113 stars 5 forks source link

Question about the SUN RGB-D metric of VA-DepthNet #12

Open jerry-ryu opened 7 months ago

jerry-ryu commented 7 months ago

I am writing to you today with a question regarding your impressive paper, "VA-DepthNet: A Variational Approach to Single Image Depth Prediction." I am particularly interested in Table 4, which presents the results on the SUN-RGB-D test set when the model is trained only on the NYU V2 training dataset.

While reading your paper, I noticed that the reported metrics for the SUN-RGB-D test set in Table 4 differ from those presented in other related works, such as the AdaBins paper by Bhat et al. (2021) and the DDP paper by Yuanfeng Ji et al. (2023). I am wondering if there might be an explanation for these discrepancies or if I have perhaps missed some important detail in my understanding.

Metrics written on VA-depthnet: image Metrics written on Adabins: image Metrics written on DDP: image

I would be very grateful if you could clarify this point for me. I am eager to learn more about your work and its impressive methodology.

Thank you for your time and consideration.

cnexah commented 6 months ago

Sorry for the late reply!

Yes, the evaluation metrics on SUN-RBGD are different. Please note we have mentioned in the paper, "In addition, we align the predictions from all the models with the ground truth by a scale and shift following Ranftl et al. (2020)."