Join the AI Career Program www.nicolai-nielsen.com/aicareer Enroll in my Technical Courses www.nicos-school.com/courses/ Camera Calibration Software and High Precision Calibration Boards camera-calibrator.com/
The focal length is read from exif info of the image (in mm) and then converted in the code to pixel values (utils has a function for it). When I print out the value of f_px, it is always None for all images. How is the model getting focal information. Also all the pictures that I use, the depth information is often wrong.
Hey Nicolai , I get an error "No module named 'depth_pro' Any idea of how to resolve this issue ? i checked the folder structure and everything seems to be exactly as you have in your video. thank you !
Thank you for the nice video. Did you also check the 3d reconstruction using open3d? Whenever I use an estimated depth from a metric MDE algorithm to generate a point cloud, I can not get a right size of objects
Most likely because of the focal length and cam params being off. Have u tested out the new one here which also gives u the focal length? Then u can set the optical center to half of img height and half of width
@@NicolaiAI Thanks for your answer. I have calibrated my camera and I have my own parameters. I have just found your new video, but I already tested Metric3D v2 and Anything Depth. But unfortunately I got need measurements in 3D and the sizes are not correct
Good one, I hace seen your old video says c++ is faster than python but why this video is using python 🤔 btw may I ask what's your gpu are using now and it's c++ and python's fps?
hi everyone, i tried using this for a project, did everything like the video. The thing is, it is way too slow. I noticed that the "model.infer(image_input, f_px=f_px)" takes about 60 seconds to run. Is there any workaround for this? Thank you for reading
@@NicolaiAI well i used fp16 precision and ,while it took some time, it was faster TLDR:the create_model_and_transforms has an argument precision) thx for the reply,awesome content
Join the AI Career Program
www.nicolai-nielsen.com/aicareer
Enroll in my Technical Courses
www.nicos-school.com/courses/
Camera Calibration Software and High Precision Calibration Boards
camera-calibrator.com/
Everytime I see the video u uploaded, I am excited cuz all these progresses are impressive and promising.
So awesome to read! Glad u like them and u made my day 💪
This is very interesting.. again thanks for the video! Looking forward to some potential content on runtime performance with this
For sure very interesting and unlocks a bunch of new opportunities in the vision space
The focal length is read from exif info of the image (in mm) and then converted in the code to pixel values (utils has a function for it). When I print out the value of f_px, it is always None for all images. How is the model getting focal information. Also all the pictures that I use, the depth information is often wrong.
Hey Nicolai ,
I get an error "No module named 'depth_pro'
Any idea of how to resolve this issue ?
i checked the folder structure and everything seems to be exactly as you have in your video.
thank you !
hey just place your main file in the ml-depth-pro-main\src\depth_pro folder
that will do
@@AmnaBano-kd7ex Thank you it worked! I appreciate it !
Hi Nicolai! I just want to ask that how can I estimate the depth of an object as well as its classification.
How much time does the code take to run fully? Does it take much time?
Thank you for the nice video. Did you also check the 3d reconstruction using open3d?
Whenever I use an estimated depth from a metric MDE algorithm to generate a point cloud, I can not get a right size of objects
Most likely because of the focal length and cam params being off. Have u tested out the new one here which also gives u the focal length? Then u can set the optical center to half of img height and half of width
@@NicolaiAI Thanks for your answer. I have calibrated my camera and I have my own parameters. I have just found your new video, but I already tested Metric3D v2 and Anything Depth. But unfortunately I got need measurements in 3D and the sizes are not correct
@NazaninPadkan the depth map is not based on your camera parameters but just an output from the ai model. So that might be one of the reasons
What is your suggestion here?
is it necessary that the image must be of high resolution?
it is posiible to implement this code in a Intel realsense camera or doing this with other versions of yolo like yolov3?
Not sure why u would use a realsense camera with this but yeah u can use any other object detection models or any algorithm to index the depth map
What's the VRAM usage of this Apple Model?
Ran out out mem a couple of times on 16gb while recoding and bunch of stuff open. Prob around 10gb
@@NicolaiAI then I've to find a new way to measure the distance with depth map
The video is pretty quiet compared to other videos and ads on RUclips, maybe bump up the volume a bit?
Yeah missed that in this one! Thanks for letting me know
No problem, thanks for the great videos!
Can it be used with IDS camera, man?
Yes as long as u can extract the frames and throw them through the model
Good one, I hace seen your old video says c++ is faster than python but why this video is using python 🤔 btw may I ask what's your gpu are using now and it's c++ and python's fps?
hi everyone, i tried using this for a project, did everything like the video. The thing is, it is way too slow. I noticed that the "model.infer(image_input, f_px=f_px)" takes about 60 seconds to run. Is there any workaround for this? Thank you for reading
Yeah , i had the same issue, if you find a solution please do let me know as well . Thanks !
Same issue for me. The command does take about 4-5 minutes on Mac Book Air, 8GB.
@@isura3147maybe try other model. I heard that depthanything is much faster
@@sandeepwadhwa531see my other comment
peace nicolai , i used the model with a mobile rtx 3050ti and it takes ages,any suggestions?
Hmm might be because of ram. But did take 20 seconds on my MacBook cpu. Will have to test it on colab or my own gpu
@@NicolaiAI well i used fp16 precision and ,while it took some time, it was faster
TLDR:the create_model_and_transforms has an argument precision)
thx for the reply,awesome content
@koufdell thanks a lot!