Great work! I was wondering if we can use it on a video? For example, run DepthAnythingV2 on every frame and then generate a consistent 3d point cloud.
Great work Bohdan, i am also working on something similar, but wanted to understand how to get rid of points b/w object which appears to be flying when the objects are at different depth
Thank you! If you mean a "gradient" point when we have a change in depth, you could use simple outliers removing filters - it will be helpful in most cases
Really great idea! I have already started preparation for this video. I want to combine depth estimation and segmentation in 2d/3d, so stay here for new videos!
@@BohdanVodianyk Thanks for your response. I tried to test on my camera with the indoor checkpoints but the error of measuring the distance is quite high. So, can we train the Depth any model on a custom dataset? or do we have any other way to reduce the error?
@@phamminhvuong8985 Of course, you could train your own model! Here is the source page of metric depth anything github.com/LiheYoung/Depth-Anything/tree/main/metric_depth It will be super useful if you want to really deeply use this technique. Also, errors in depth could be very sensible to the camera parameters and light conditions, so be sure you have enough light as in my video for example
thanks for the great work Bohdan!
Thanks, brother
Been waiting on this!
hehehe, thank you! I am preparing many new cool things, so in the future check out many great things here
Great work! I was wondering if we can use it on a video? For example, run DepthAnythingV2 on every frame and then generate a consistent 3d point cloud.
You can go from RGB-D video to point cloud with e.g. SFM or SLAM algorithms.
Great work Bohdan, i am also working on something similar, but wanted to understand how to get rid of points b/w object which appears to be flying when the objects are at different depth
Thank you! If you mean a "gradient" point when we have a change in depth, you could use simple outliers removing filters - it will be helpful in most cases
How about real time semantic segmentation with pcd for next video topic?
Really great idea! I have already started preparation for this video. I want to combine depth estimation and segmentation in 2d/3d, so stay here for new videos!
which version of depth anything do you apply in your code to reconstruct the point cloud?
For metric depth estimation, I have used pre-trained models (indoor and outdoor) that are based on Depth Anything v1
@@BohdanVodianyk Thanks for your response. I tried to test on my camera with the indoor checkpoints but the error of measuring the distance is quite high. So, can we train the Depth any model on a custom dataset? or do we have any other way to reduce the error?
@@phamminhvuong8985i have the same question
@@phamminhvuong8985 Of course, you could train your own model! Here is the source page of metric depth anything github.com/LiheYoung/Depth-Anything/tree/main/metric_depth It will be super useful if you want to really deeply use this technique. Also, errors in depth could be very sensible to the camera parameters and light conditions, so be sure you have enough light as in my video for example
@@BohdanVodianyk Coud you share a tutorial video about training depthany model on custom data?