Reversed Depth from Midas

[penguinflys]

FSGS/train.py

Lines 121 to 131 in a536a64

	render_pkg_pseudo = render(pseudo_cam, gaussians, pipe, background)
	rendered_depth_pseudo = render_pkg_pseudo["depth"][0]
	midas_depth_pseudo = estimate_depth(render_pkg_pseudo["render"], mode='train')
	rendered_depth_pseudo = rendered_depth_pseudo.reshape(-1, 1)
	midas_depth_pseudo = midas_depth_pseudo.reshape(-1, 1)
	depth_loss_pseudo = (1 - pearson_corrcoef(rendered_depth_pseudo, -midas_depth_pseudo)).mean()
	if torch.isnan(depth_loss_pseudo).sum() == 0:
	loss_scale = min((iteration - args.start_sample_pseudo) / 500., 1)
	loss += loss_scale * args.depth_pseudo_weight * depth_loss_pseudo

From above code, I see depth are transformed to the opposite number, but i guess that is not what reverse means, shouldn't it be 1/midas_depth_pseudo instead -midas_depth_pseudo?

Reference: isl-org/MiDaS#261

Best Regards

[penguinflys]

As the papar only utilize relative depth, how to convert it to the depth seems not a big deal, i tried to use the 1/d way to produce real depth, but the rendering result gets worse. Perhaps hyperparameter should also be tuned.

more info can be found in #1

Anyways, it seems not to be a big deal.