hey im a 3rd year student. I'm too planning to do a 3d reconstruction using a 2d image for my 4th year project( i have a year more but i want to start now). I have no idea where to start. is there any way i can reach out
Thanks a lot. The accuracy for real world distances is low and u should probably not expect more than down to cm accuracy. And that's if u have some good mapping from relative to absolute depths
Thanks a lot for watching! Means a ton. Yeah u can fine tune models on ur own dataset it’s just pretty hard to get. You should def check out the Midas GitHub repo for more info. All of it is in there
Hi, nice video BTW, i have a question (maybe dumb but here it comes) If you run this MiDas implementation on an image, and you go the output (no the image, the prediction matrix) and you want to get the depth of the pixel [20][20] of an image. Suppuse that value is 20, what does this 20 means? The pixel [20[20] is 20m (or centimiteres or whaT?) far away from the camera? Or to know that distance you have to do something more? Thanks in advance.
Thanks a lot for watching! All the depths are relative to the camera and u will have to make a mapping urself to get absolute distances. It could be that u have some references in the scene
@@NicolaiAI THanks for your response. So, if i have understood, lets say i know the real distance of pixel A and pixel B. Knowing the relative distance captured with MiDaS, there would be a relationship as a simple rule of three? And this would be true for every images of the same camera?
Thanks a lot, to get the relative distances u can just project the points out in the world and calculate the distance. But if u want the absolute distances u will have to have some references that u can use for mapping ur depth values to the real world
Thank you so much ...this is very interesting ,actually I want to build a robot car based on depth estimation for successful navigation and I wanna train a model and make the car go to the deepest point ... but the problem that Midas is trained on 10 datasets seems impossible to train on this verity of datasets do you have any suggestion on how to caluclate the deepest point or a good model with high prediction and can I train it and understand it by myself ..thanks you so much again for helping me in my school project.
Man your works are amazing I really appreciate. However, I don't understand that how you are aware of those technologies? I just heard by you. I mean how I can learn them own my own?
Thanks Nicolai for your great videos, but do you know how to convert depth_map = prediction.cpu().numpy() to meters/cm/mm? what exactly the output is? I modified this code to detect objects with Yolo3 but I want to convert the mean object value to the meter.
![Blox Fruits Dragon Rework Update [Full Stream]](http://i.ytimg.com/vi/EqDAp8udhm0/mqdefault.jpg)
Join My AI Career Program
www.nicolai-nielsen.com/aicareer
Enroll in My School and Technical Courses
www.nicos-school.com
Finally a machine learning project I'm able to follow along with! I modified the DPT code so now I can run semantic segmentation with my webcam. :D
Thanks a lot for watching! Would u mind to share the code with me? Would love to make a video about it. thecodinglibraryyt@gmail.com
You are an excellent professor!
Thanks for watching!
My graduation project is a 3D reconstruction system. Your videos are so helpful to me. Thank you!
Cool! Thanks for watching, really glad that I can help
hey im a 3rd year student. I'm too planning to do a 3d reconstruction using a 2d image for my 4th year project( i have a year more but i want to start now). I have no idea where to start. is there any way i can reach out
Please mention the versions of all the dependencies
Hi thank you for the great video. Can you tell how to convert relative depth maps to absolute depth maps?
Great video ! Is there any way to convert the depth you capture from the model into Meter ?
is this model better than the old monocular depth estimation?
Yeah way better actually
@@NicolaiAI do you have the implementation of the model from scratch?
Cool Video (y). Any Idea on the accuracy of such an approach. Can I relay on translating the pixel density to mm or cm? Thanks
Thanks a lot. The accuracy for real world distances is low and u should probably not expect more than down to cm accuracy. And that's if u have some good mapping from relative to absolute depths
thank you so much it's really interesting content
I have a question:
is it not possible to fine tune Midas on a database? and how to do it?
Thanks a lot for watching! Means a ton.
Yeah u can fine tune models on ur own dataset it’s just pretty hard to get. You should def check out the Midas GitHub repo for more info. All of it is in there
Hi sir, can this be used for measuring the object distance from the camera??
We could map the coulrs to depth approximations
@@swastikmohapatra4665 can you please suggest a code
So that step with interpolation is a reverse logic of above input transformations?
nice depth map
It's really amazing!
Nice video. Can this be use for object dimensions estimation?
Hi, nice video BTW, i have a question (maybe dumb but here it comes)
If you run this MiDas implementation on an image, and you go the output (no the image, the prediction matrix) and you want to get the depth of the pixel [20][20] of an image. Suppuse that value is 20, what does this 20 means? The pixel [20[20] is 20m (or centimiteres or whaT?) far away from the camera? Or to know that distance you have to do something more?
Thanks in advance.
Thanks a lot for watching! All the depths are relative to the camera and u will have to make a mapping urself to get absolute distances. It could be that u have some references in the scene
@@NicolaiAI THanks for your response.
So, if i have understood, lets say i know the real distance of pixel A and pixel B. Knowing the relative distance captured with MiDaS, there would be a relationship as a simple rule of three?
And this would be true for every images of the same camera?
Hello,blogger, are there more papers doing supervised monocular depth estimation now?
Hello, how to use this model to get the absolute distance? What code needs to be modified?
Well, it seems to be quite slow. Are you using GPU in this video or not ?
Yes it's running on a gtx 1060. So not a great GPU but if u take the smaller models u will get faster inference
Interesting video :) Do you know how to get the relative distances between the objects?
Thanks a lot, to get the relative distances u can just project the points out in the world and calculate the distance. But if u want the absolute distances u will have to have some references that u can use for mapping ur depth values to the real world
thanks for answering, how could I project the points and calculate the distance. I am struggling with that. I will appreciate any help 😊
Very good 😃 🇧🇷👊
Thanks for watching!
hey, do you know how to use video instead of images as input for midas?
Videos is basically just images. U would just read in the path to ur video instead of the webcam.
@@NicolaiAI I downloaded MiDaS and I can set input images by putting them on the input folder. But videos aren't read
HELLO
Can I train this nerual network for single-shot structured light depth sensgin system (speckle IR pattern just like iPhone system)?
what's the deference between depth map and the grayscale one?
where can I get the above code?
Thank you so much ...this is very interesting ,actually I want to build a robot car based on depth estimation for successful navigation and I wanna train a model and make the car go to the deepest point ... but the problem that Midas is trained on 10 datasets seems impossible to train on this verity of datasets do you have any suggestion on how to caluclate the deepest point or a good model with high prediction and can I train it and understand it by myself ..thanks you so much again for helping me in my school project.
Did you find the solution?
Can we do this on a video input other than webcam?
Yes u can do that by replacing the camera index with the path to the video file
Thank you! Can we calculate distance from this?
@@sadiachowdhury3630 yes the Midas doesn't care what kind of input you use, as long as it is an image or imageseries
Does it require stereo image
Nope this is just for a single camera!
Hi, I am very new to computer vision , I am running this code in jupyter notebook but it is not showing any output , although it ran successfully
Apparently 2023 is the first year people started referring to CNNs as 'old school' and 'old fashioned'.
CNNs are still far ahead of transformer on inference speed
Can we train MiDas with our own data?
Can someone explain me how to get the weights of the model???
please we want a neural network in the case of stereovison too
Please share the repo link next time. We really appreciate that, I'm sure you'll get more likes if you do
Man your works are amazing I really appreciate. However, I don't understand that how you are aware of those technologies? I just heard by you. I mean how I can learn them own my own?
Thanks a lot for watching and the support! I just look up the latest and most cited papers within different subjects
@@NicolaiAIDo you mean academic papers?
@@johnconnor8717 yeye
@@NicolaiAI I see thanks for informing :)
and this is all with a regular camera right?
Yup works on every camera
Thanks Nicolai for your great videos, but do you know how to convert depth_map = prediction.cpu().numpy() to meters/cm/mm? what exactly the output is? I modified this code to detect objects with Yolo3 but I want to convert the mean object value to the meter.
Thanks for watching! This is all relative normalized distances so I will have to do a mapping urself to another unit in ur environment
Is there existing onnx of newest midas model ? 😇
Look at their git lol
@@kruemelkevin allready did, i had one for the last version, not for this one. Thx for the tips btw
Does anyone know if this depth map is absolute or relative?
nevermind it is relative, is there a way to get this in absolute depth ?
@@aatkarelse8218 yeah it's the relative distance. U need some mapping with references if u want the absolute distances
@@NicolaiAI yes, i was afraid it would be something like that, thx.
Where's the discord server link?
can you please send me this model from scratch?