search

princeton-vl / DROID-SLAM

BSD 3-Clause "New" or "Revised" License
1.75k stars 295 forks source link

Scale problem of droid-slam #85

Open gyes00205 opened 1 year ago

gyes00205 commented 1 year ago

Hi authors, Thanks for your amazing work. I have a question about scale of camera pose. I use ZED camera to collect data and compare camra pose of ZED and droid-slam. Blue dot means camera pose of ZED and red dot means camera pose of droid-slam.

image

I find some scale problems of droid-slam. Could you tell me how to solve the scale problem of camera pose? Thanks.

kg0221 commented 10 months ago

Were you able to solve this? I'm seeing something similar. I expect good scale because I seed the initial samples with actual measured poses. But the scale seems to shrink after it gets beyond these initial values.

gyes00205 commented 9 months ago

Hi @kg0221, you can refer to this reply .

kg0221 commented 9 months ago

Thanks @gyes00205. But I'm not using a stereo camera. So I don't think that solution works for me. However, I did find that the training dataloader scales the world coordinates by 1/5. And the results are not scaled back up afterwords. I'm not sure why this is. But by scaling my inputs by 1/5 and my outputs back up by 5 afterwords, I was able to get my expected results.

lbg030 commented 4 months ago

Thanks @gyes00205. But I'm not using a stereo camera. So I don't think that solution works for me. However, I did find that the training dataloader scales the world coordinates by 1/5. And the results are not scaled back up afterwords. I'm not sure why this is. But by scaling my inputs by 1/5 and my outputs back up by 5 afterwords, I was able to get my expected results.

Hi, I encounter with same problem. I want to set GT pose, only evaluating 3D reconstruction. Can you explain how to do ? thank you so much

Promethe-us commented 3 weeks ago

Thanks @gyes00205. But I'm not using a stereo camera. So I don't think that solution works for me. However, I did find that the training dataloader scales the world coordinates by 1/5. And the results are not scaled back up afterwords. I'm not sure why this is. But by scaling my inputs by 1/5 and my outputs back up by 5 afterwords, I was able to get my expected results.

Hello, I meet the save problem, have you solved it? I set gt w2c for each iter, but depth tend to be worse.