fdarmon / wild_deep_mvs Goto Github PK

Hi there,

Thanks for the great work!

I'm trying to re-compute the metrics reported in Table 4 of your paper (prec., rec., f-score on YFCC evaluation) but there doesn't appear to be code in your repo for doing this. I have my own function for computing the metrics, but am a bit confused about setting a threshold.

First, do you have the thresholds you used for each YFCC scene stored anywhere? I tried using the values stored in the text files in yfcc_data/gt_resolution, but using your provided trained models with this threshold doesn't give me the same results you report in Table 4.

Second, is Eq. 7 in the paper correct? I understand the goal with Eq. 7, but shouldn't the distance argument be ||K^-1 (D(p)p) - K^-1 (D(p')p')||. This way, you find the median distance in scene space between back-projected points 2 pixels away from each other. Is this perhaps what you used or did you use Eq. 7 as is?

Thanks!
Alex

hello, I find that the download link of BlendedMVS has expired, can you provide a new one? thanks very much

Hi, thanks for a good work.
Could you provide your exact parameters for generating gt through fusing depth maps from IMC? I was fusing these scenes with parameters specified in paper "reprojection error below half a pixel and depth error below 1%", but the fusing result is not as good as the gt you provided. Thanks!

Hi, thank you for sharing both the paper and the code. I'm working on something similar, so I was very happy to read your results.

I would like to ask if you ever considered urban datasets in your evaluation, especially multi-camera datasets such as nuScenes or DDAD. I'm asking this for two main reasons:

1. Internet data are surely "in the wild", but they mostly focus on a single giant object (i.e. a building) in the image, for which a lot of diverse views can be captured. On the other hand, urban data have no clear subject, a lot of dynamic objects and several textureless areas to deal with, which is definitely an even harder test for MVS networks.
2. MVS networks cannot be trained in a supervised way on urban data, therefore your insights on unsupervised methods might be interesting to be validated also on these kind of data.

What are your thoughts on this?