Question about leaf nodes unchanged during optimization

[Yang-Xijie]

In paper 5.1 Optimizing the Hierarchy, you mention "we do not change the leaf nodes during optimization." However, I cannot find the according code in train_post.py.

---

In train_post.py, threshold, the third parameter of expand_to_size(), is set to [0.005, 0.1].

hierarchical-3d-gaussians/train_post.py

Lines 66 to 67 in 85777b1

	limmax = 0.1
	limmin = 0.005

hierarchical-3d-gaussians/train_post.py

Lines 74 to 75 in 85777b1

	limit = math.pow(2, sample * (math.log2(limmax) - math.log2(limmin)) + math.log2(limmin))
	scale = 1

hierarchical-3d-gaussians/train_post.py

Lines 91 to 99 in 85777b1

	to_render = expand_to_size(
	gaussians.nodes,
	gaussians.boxes,
	limit * scale,
	viewpoint_cam.camera_center,
	torch.zeros((3)),
	render_indices,
	parent_indices,
	nodes_for_render_indices)

I speculate that limmin helps keep the leaf nodes unchanged. Since the larger the limmin, the lower possibility the leaf nodes will be selected. Is this correct?

[Yang-Xijie]

I find the following code keeps anchors unchanged.

hierarchical-3d-gaussians/train_post.py

Lines 164 to 181 in 85777b1

	# Optimizer step
	if iteration < opt.iterations:
	if gaussians._xyz.grad != None:
	if gaussians.skybox_points != 0 and gaussians.skybox_locked: #No post-opt for skybox
	gaussians._xyz.grad[-gaussians.skybox_points:, :] = 0
	gaussians._rotation.grad[-gaussians.skybox_points:, :] = 0
	gaussians._features_dc.grad[-gaussians.skybox_points:, :, :] = 0
	gaussians._features_rest.grad[-gaussians.skybox_points:, :, :] = 0
	gaussians._opacity.grad[-gaussians.skybox_points:, :] = 0
	gaussians._scaling.grad[-gaussians.skybox_points:, :] = 0
	gaussians._xyz.grad[gaussians.anchors, :] = 0
	gaussians._rotation.grad[gaussians.anchors, :] = 0
	gaussians._features_dc.grad[gaussians.anchors, :, :] = 0
	gaussians._features_rest.grad[gaussians.anchors, :, :] = 0
	gaussians._opacity.grad[gaussians.anchors, :] = 0
	gaussians._scaling.grad[gaussians.anchors, :] = 0

However, when I use my own dataset, it seems that nothing exists in anchors.bin: the content of anchors.bin is 0x00000000. In this case, the leaf nodes are still optimized.

[Yang-Xijie]

related PR: graphdeco-inria/gaussian-hierarchy#1