我要如何实现 about deepmatchvo HOT 14 CLOSED

@XAB745220zt Hi, do you mean estimate depths using your own data? You can just resize your images to the input size and feed them to the network (refer to demo.py). You may also need to re-train the network to get better results since these cameras may have different intrinsic parameters.

from deepmatchvo.

Yes, I want to use the camera to collect the video, and then measure the distance of the object. The problem is this. My picture is 640x480. I detected the center coordinate of the object. I want to get the distance through this coordinate, so I can measure the world coordinate X, Y. ,Z

from deepmatchvo.

But I am afraid of resize, after the problem, I want the output to be the same size as the original image.

from deepmatchvo.

@hlzz HI ， I don't understand how the following is calculated. I want to calculate the world coordinates X, Y, Z of a certain coordinate. How to proceed?

def normalize_depth_for_display(depth, pc=95, crop_percent=0, normalizer=None, cmap='gray'):
# convert to disparity
depth = 1./(depth + 1e-6)
if normalizer is not None:
depth = depth/normalizer
else:
depth = depth/(np.percentile(depth, pc) + 1e-6)
depth = np.clip(depth, 0, 1)
depth = gray2rgb(depth, cmap=cmap)
keep_H = int(depth.shape[0] * (1-crop_percent))
depth = depth[:keep_H]
return depth

from deepmatchvo.

The depth value you obtain using this method is up to scale, which means you cannot get absolute coordination that is aligned with the real world space.

from deepmatchvo.

@hlzz 呃呃呃呃，我好难过，我是否可以获取视差?

from deepmatchvo.

@hlzz Do you have any good suggestions for me to think this

from deepmatchvo.

Thanks for your interest. I may not fully understand you but this work may not fulfill your need. Can you elaborate what you want to do?

from deepmatchvo.

@hlzz Thank you for being so patient to listen to me as a newcomer. I want to do something like this. I can detect the object through yolo, then get the center coordinates of the frame, and get the world coordinates X, Y, Z of the object through stereo vision. I used opencv. I tried the effect is not ideal, it may be that I will not optimize the stereo matching algorithm of opencv

from deepmatchvo.

Therefore, I turned to deep learning and obtained parallax. I don’t know what depth estimation is, and I don’t know the difference between it and stereo matching in deep learning.

from deepmatchvo.

I guess you want to detect an object in the image and get its world coordinate? Then you need to first define a coordination system, right? If you want to do it with monocular depth estimation (like this work). The depth value you obtain cannot be aligned with the real world space that we live in. So let’s say you have two cameras with known baseline (the distance between them), it is feasible to get its world coordinate, by stereo matching. This is the difference between monocular depth estimation and stereo vision. The former needs only one image, the latter needs image pairs.

from deepmatchvo.

@hlzz Yes, I have a binocular camera with a parallel model, but I don't know where to start, i seems to try all the methods.，Can you tell me how to do it?

from deepmatchvo.

You can start with opencv. There are plenty of tutorials for stereo matching using opencv

from deepmatchvo.

@hlzz I tried BM, SGBM, but the effect is not good.

from deepmatchvo.