🐛 Bug: post_process_mesh fails with small number of clusters (invalid indexing)

[filippocastelli]

Description

In utils/mesh_utils.py, the function post_process_mesh() fails when the number of detected triangle clusters is smaller than cluster_to_keep.

Specifically, this line:

n_cluster = np.sort(cluster_n_triangles.copy())[-cluster_to_keep]

can raise an IndexError due to invalid negative indexing when:

cluster_to_keep > len(cluster_n_triangles)

---

Location

2d-gaussian-splatting/utils/mesh_utils.py

Line 35 in 335ad61

n_cluster = np.sort(cluster_n_triangles.copy())[-cluster_to_keep]

---

Why this happens

• cluster_n_triangles contains one entry per detected cluster
• If the mesh has few connected components (e.g. < 1000 clusters), then:

np.sort(cluster_n_triangles)[-cluster_to_keep]

tries to access an index outside bounds → crash

---

Expected behavior

The function should gracefully handle cases where the number of clusters is smaller than cluster_to_keep.

---

Proposed fix

Clamp cluster_to_keep to the available number of clusters before indexing:

num_clusters = len(cluster_n_triangles)
clusters_to_keep = min(cluster_to_keep, num_clusters)

n_cluster = np.sort(cluster_n_triangles.copy())[-clusters_to_keep]

Also, renaming cluster_to_keep → clusters_to_keep would better reflect its meaning.

---

Optional improvement

Avoid sorting entirely (more efficient):

clusters_to_keep = min(cluster_to_keep, len(cluster_n_triangles))
n_cluster = np.partition(cluster_n_triangles, -clusters_to_keep)[-clusters_to_keep]
n_cluster = max(n_cluster, 50)

---

Additional notes

• This bug appears frequently when working with small or clean meshes (few connected components)
• The current default (cluster_to_keep=1000) makes the issue very likely in typical cases

---

Environment

• Repo version: 335ad61
• Python: (3.12) (using 3.12 requires a fix in diff-surfel-rasterization)
• Open3D: (0.19.0)