search

sshan-zhao / ACMNet

Adaptive Context-Aware Multi-Modal Network for Depth Completion
64 stars 11 forks source link

Invariance to Sparse Sample #14

Open TouqeerAhmad opened 12 months ago

TouqeerAhmad commented 12 months ago

I have tried running the evaluation script with the default number of samples i.e., 10000 5000 2500 at three scales and have also tried with lower numbers e.g., 1000 500 250 and even 100 50 25. I do not see any significant change in the visual results and the numerical metrics also seem very close. While I understand the method is somewhat invariant to sparsity, I was expecting performance to degrade severely especially with 100 50 25 sampling. Can you please advise?

TouqeerAhmad commented 12 months ago

I was trying to read the paper and I think I am a bit more confused, may be author or any of the participants in issues can help me. In section IV-A, paper says: "then sample 500 sparse LiDAR points from the provided dense depth map randomly as the sparse depth data."

then in section IV-B it says: "For KITTI dataset, we build the graphs at three scales with 10000, 5000, and 2500 observed pixels randomly sampled from the downsamples sparse depth maps"

How 10000 pixels can be observed from only 500 sparse samples?

sshan-zhao commented 9 months ago

I have tried running the evaluation script with the default number of samples i.e., 10000 5000 2500 at three scales and have also tried with lower numbers e.g., 1000 500 250 and even 100 50 25. I do not see any significant change in the visual results and the numerical metrics also seem very close. While I understand the method is somewhat invariant to sparsity, I was expecting performance to degrade severely especially with 100 50 25 sampling. Can you please advise?

Hi, the parameter nsamples does not indicate the number of sparse points with depth values, while representing the number of points for building graph. The tables Tab III and Tab VI show some analysis.

sshan-zhao commented 9 months ago

I was trying to read the paper and I think I am a bit more confused, may be author or any of the participants in issues can help me. In section IV-A, paper says: "then sample 500 sparse LiDAR points from the provided dense depth map randomly as the sparse depth data."

then in section IV-B it says: "For KITTI dataset, we build the graphs at three scales with 10000, 5000, and 2500 observed pixels randomly sampled from the downsamples sparse depth maps"

How 10000 pixels can be observed from only 500 sparse samples?

"In section IV-A, paper says: "then sample 500 sparse LiDAR points from the provided dense depth map randomly as the sparse depth data."" This expression is for NYU v2, not KITTI. Maybe the word "LiDAR" confuses you. Sorry for that.