the label_scale_factor mean depth scale or the pixel scale of the camera ? about metric3d HOT 13 CLOSED

what is the depth range when you train the depth network? we add the depth network to our model and freeze, the whole model will be difficult to train if without normalization the depth map

The bin-based prediction is ranged in (0.3, 150), defined here. Afterwards a refine module is imposed, so theoratically there is no such range for the final output. In the meanwhile sky is defined to be 180m. Somehow a small propotion of backgrounds might be even "farther" than sky, under this setting.
did you normalization the depth to (0,1)? when you train the depth network to calculate the depth loss

Bin-based prediction should be regarded as such normalization. It describes a probability distribution upon pre-defined depth bins.

from metric3d.

the label_scale_factor mean depth scale or the pixel scale of the camera? or any others?

It is an image-level scale factor from the raw camera space to the canonical space.

from metric3d.

we add the depth module to our model, but the output from depth module is so small , the max is 5.04 (intrinsic[1000,1000,160,80])

from metric3d.

we add the depth module to our model, but the output from depth module is so small , the max is 5.04 (intrinsic[1000,1000,160,80])

The image size looks fairly small. You could set the crop_size in the config file to a smaller one.

from metric3d.

I do it, but there are problems. I set the img_size=(480, 854), focal_length=1000.0, (intrinsic[1000,1000,427,240]) the max value in the depth maps is tensor(24.0547, device='cuda:0'), but tested same image in your test. sh the max the is 260.7979. ( the setting is same) I can't understand.

from metric3d.

I do it, but there are problems. I set the img_size=(480, 854), focal_length=1000.0, (intrinsic[1000,1000,427,240]) the max value in the depth maps is tensor(24.0547, device='cuda:0'), but tested same image in your test. sh the max the is 260.7979. ( the setting is same) I can't understand.

Oh I think we should never change image_size, which means the image size during training (for the canonical space camera). We shall change crop_size instead.

from metric3d.

I see. I change the crop_size to 480, 854

from metric3d.

I do it, but there are problems. I set the img_size=(480, 854), focal_length=1000.0, (intrinsic[1000,1000,427,240]) the max value in the depth maps is tensor(24.0547, device='cuda:0'), but tested same image in your test. sh the max the is 260.7979. ( the setting is same) I can't understand.

Here do you mean that you have a groundtruth "depth map"? If it is an outdoor image, the model is likely to predict very large values for very far backgrounds.

from metric3d.

no without gt ,we predict a drive image

from metric3d.

no without gt ,we predict a drive image

If the pixel with 260-meter depth lies in very ranged regions, it could be possible.

from metric3d.

what is the depth range when you train the depth network? we add the depth network to our model and freeze, the whole model will be difficult to train if without normalization the depth map

from metric3d.

what is the depth range when you train the depth network? we add the depth network to our model and freeze, the whole model will be difficult to train if without normalization the depth map

The bin-based prediction is ranged in (0.3, 150), defined here. Afterwards a refine module is imposed, so theoratically there is no such range for the final output. In the meanwhile sky is defined to be 180m. Somehow a small propotion of backgrounds might be even "farther" than sky, under this setting.

from metric3d.

what is the depth range when you train the depth network? we add the depth network to our model and freeze, the whole model will be difficult to train if without normalization the depth map

The bin-based prediction is ranged in (0.3, 150), defined here. Afterwards a refine module is imposed, so theoratically there is no such range for the final output. In the meanwhile sky is defined to be 180m. Somehow a small propotion of backgrounds might be even "farther" than sky, under this setting.
did you normalization the depth to (0,1)? when you train the depth network to calculate the depth loss

from metric3d.