search

facebookresearch / consistent_depth

We estimate dense, flicker-free, geometrically consistent depth from monocular video, for example hand-held cell phone video.
MIT License
1.62k stars 235 forks source link

Customized run result not reasonable #6

Closed lawy623 closed 4 years ago

lawy623 commented 4 years ago

Thanks for sharing the code!

I run python main.py --video_file $video_file_path --path $output_path --make_video, with no calibrate camera. But the output results do not look reasonable.

Any suggestion for this case? I see the finetuning makes disparity-0.1 and loss-0.5. Must I provided the calibrated intrinsics to run it well?

image

roxanneluo commented 4 years ago

Can you share your video? Larger baselines typically improves the results. Did you inspect the clomp reconstruction? Did the camera register correctly?

lawy623 commented 4 years ago

I upload the video I filmed by iphone here: https://drive.google.com/file/d/1-iDkf5KWkxfdSp5hx2TqYhrc0q_6W7cj/view?usp=sharing

I inspect the depth_colmap_dense folder under the result directory and find this. Probably the reconstruction is not correct? WeChatWorkScreenshot_c75335e4-5485-432e-a942-49075f7f0908

I did not get into the code yet but I run with the default setting and change nothing except for setting the video_path. I am not sure how should I make the reconstruction more reliable.

lawy623 commented 4 years ago

I change the video input and find that the result gets improved. I think it all about the colmap reconstruction but it has very hard restriction. Thanks for the sharing~