Output of grayscale depthmaps

[herrartist]

How can I output grayscale depthmaps (and nothing else)?

[Dhruva-Storz]

Not sure if depth pro has an option for this, but the output of the model is a metric depth array. You need to normalize the depth values between 0 and 255 to get a traditional 8 bit greyscale image (and possibly invert it), then save the file as a PNG or whatever format.

[hradec]

You can run depth_pro and save a floating point tif file using something like this snippet:

        import depth_pro
        import torch, cv2

        def get_torch_device() -> torch.device:
            """Get the Torch device."""
            device = torch.device("cpu")
            if torch.cuda.is_available():
                # use gpu if it has more than 7GB free
                if torch.cuda.get_device_properties(0).total_memory/1024/1024/1024 > 7:
                    device = torch.device("cuda:0")
            elif torch.backends.mps.is_available():
                device = torch.device("mps")
            return device

        # Load model and preprocessing transform
        model, transform = depth_pro.create_model_and_transforms(
            device=get_torch_device(),
            precision=torch.half,
        )
        self.progress_callback()
        model.eval()

        # Load and preprocess an image.
        image, _, f_px = depth_pro.load_rgb("your_rgb_image.jpg")
        image = transform(image)

        # Run inference.
        prediction = model.infer(image, f_px=f_px)
        depth = prediction["depth"].detach().cpu().numpy().squeeze()
        cv2.imwrite(f"your_rgb_image_depthPro.tif", depth)

A floating point tif will give you the full range and precision of the calculated depth, including values above 1 meter.

You can also use the depth variable straight from it just like a normal cv2.imgread() object. You don't have to save to tif and read back with cv2.imgread().

But be careful when loading up the tif file in image editors like photoshop or gimp, since all of then will apply some sort of color conversion instead of display the actual floating point values.

You can use this script to display the tif images properly, and sample the values using the mouse:
https://gist.github.com/hradec/adb03ea9fd02ee2a8cde00a69fb26f62