PointCloudEnhancement – Evaluation Metrics

This repository provides a complete Python implementation for evaluating the quality of 3D point-clouds, especially for point-cloud enhancement, reconstruction, and completion tasks.
The metrics implemented here cover accuracy, completeness, normal consistency, F-score, and two types of Chamfer distance.

The core evaluation logic is implemented in
metrics.py, which has been fully documented and explained in Explanation.md.

---

📦 Features

The toolkit computes the following metrics between a predicted point cloud and a ground-truth point cloud:

Geometry-based metrics

• Accuracy (CD_Acc)
  Mean nearest-neighbor distance from predicted → ground truth.
• Completeness (CD_Comp)
  Mean nearest-neighbor distance from ground truth → predicted.
• Symmetric Chamfer distances
  • chamfer-L1 — L1 Chamfer distance (sum of mean distances in each direction)
  • chamfer-L2 — L2 Chamfer distance (sum of squared mean distances)
  • chamferL2_old — legacy symmetric Chamfer = 0.5 × (accuracy + completeness)

Normal-based metrics

• Normal Accuracy (N_Acc)
  Cosine similarity between predicted normals and nearest neighbors in GT.
• Normal Completeness (N_Comp)
  Cosine similarity from GT → predicted.
• Normal Correctness (normals)
  Mean of the two normal terms (0.5 × N_Acc + 0.5 × N_Comp)

F-Score family (threshold-based metrics)

Computed at configurable distance thresholds (default: 5, 10, 20 units):

For each τ:

• Precision Pτ = ratio of predicted points within τ of GT
• Recall Rτ = ratio of GT points within τ of prediction
• Fτ = 2PR/(P+R)

Example output keys: P_5, R_5, F_5 P_10, R_10, F_10 P_20, R_20, F_20

---

🚀 Usage

Function call

The core entry point is:

from metrics import eval_pointcloud

results = eval_pointcloud(
    pre_mesh_ply="pred.ply",
    gt_mesh_ply="gt.ply",
    samplepoint=,
    eval_type=""
)