---

Implementation of the Paper

Textured Mesh Quality Assessment using Geometry and Color Field Similarity

---

1. Single Mesh Quality Evaluation

1.1 Command Line Usage

python fmqm_single_mesh_eval.py \
    --ref_obj_path /path/to/reference.obj \
    --ref_tex_path /path/to/reference_texture.png \
    --dis_obj_path /path/to/distorted.obj \
    --dis_tex_path /path/to/distorted_texture.png

1.2 Required Parameters

Parameter	Required	Description
--ref_obj_path	Yes	Path to the reference mesh (.obj)
--ref_tex_path	Yes	Path to the reference texture (.png)
--dis_obj_path	Yes	Path to the distorted mesh (.obj)
--dis_tex_path	Yes	Path to the distorted texture (.png)

1.3 Hardcoded Hyperparameters

# Sampling configuration
noise_level = 5
point_number = 200000
lp_number = 250
sample_mode = "fps"

# Feature extraction thresholds
tangentThreLevel = 5
normalThreLevel = 5
tangentPlaneLevel = 10
radius_level = 1.25

1.4 Output

Prints a 5-dimensional FMQM score to the console:

FMQM Score: [f1, f2, f3, f4, f5]

---

2. Dataset-Level Evaluation on TSMD, SJTU-MQA, YANA

2.1 Step 1: Generate Path Configuration Files

Run the following script to generate CSV files listing reference and distorted mesh paths:

python generate_path.py \
  --tsmd /path/to/TSMD/dataset \
  --sjtumqa /path/to/SJTU-MQA/dataset \
  --yana /path/to/YANA/dataset

This will generate the following files in the datasetInfo/ folder:

datasetInfo/
├── pathsTSMD.csv
├── pathsSJTUMQA.csv
└── pathsYANA.csv

Each CSV includes: modelRefName, modelDisName, ref_obj_path, ref_tex_path, dis_obj_path, dis_tex_path.

---

2.2 Step 2: Sample Geometry and Color Fields (SDF + NCF)

2.2.1 Basic Usage

Run the default configuration aligned with the paper:

python fmqm_sample_dataset.py --dataset [tsmd|sjtumqa|yana]

2.2.2 Full Parameter Options

python fmqm_sample_dataset.py \
    --dataset [tsmd|sjtumqa|yana] \
    --noise_level 5 \
    --point_number 200000 \
    --lp_number 250 \
    --sample_mode [fps|random]

2.2.3 Parameter Descriptions

Parameter	Type	Default	Description
--dataset	string	—	Dataset name (tsmd, sjtumqa, yana)
--noise_level	float	5	Noise variance = (value × 0.01)²
--point_number	int	200000	Number of sample points per mesh
--lp_number	int	250	Number of local patches
--sample_mode	string	fps	Sampling mode: fps or random

2.2.4 Output Directory Structure

sampleResults/
└── {dataset}/
    ├── SDFU{noise}_{lp_number}_{points}/
    │   ├── reference/
    │   │   ├── {model_name}.mat
    │   │   └── {model_name}_bbox.mat
    │   └── distorted/
    │       └── {model_name}.mat
    └── topology/
        └── topology_{lp_number}_{sample_mode}/
            └── {model_name}.npz

2.2.5 Sample File Format

Reference Samples (reference/*.mat)

{
    'SamplePointXYZ':       [N, 3],
    'NearestPointXYZ':      [N, 3],
    'NCF':                  [N, 3],
    'NearestPointNormal':   [N, 3],
    'SDF':                  [N, 1],
    'LocalPatchIdx':        [N, 1]
}

Bounding Box (reference/*_bbox.mat)

{
    'box_center':   [lp_number, 3],
    'box_diagonal': [lp_number, 1]
}

Distorted Samples (distorted/*.mat)

• Same structure as reference samples.

Topology Data (topology_{lp_number}_{sample_mode}/*.npz)

{
    "local_0": { 'idx': int, 'v': [N], 'f': [M], 'rings': int },
    ...
}

Note: Can be reused across sampling runs if lp_number and sample_mode are constant.

---

2.3 Step 3: Compute FMQM Scores

2.3.1 Command Line

python fmqm_calculate_scores.py \
    --dataset=[tsmd|sjtumqa|yana] \
    --sample_dir_name=SDFU5.00_250_200000

2.3.2 Parameters

Parameter	Type	Description
--dataset	string	Dataset name (tsmd, sjtumqa, yana)
--sample_dir_name	string	Sampling folder name (e.g., SDFU5.00_250_200000)

2.3.3 Hardcoded Hyperparameters

tangentThreLevel = 5
normalThreLevel = 5
radius_level = 1.25
tangentPlaneLevel = 10
numberFeatures = 5  # 0: geoSSIM, 1: geoGradientSSIM, 2: colorSSIM, 3: colorGradientSSIM, 4: FMQM

2.3.4 Output

Results saved in:

sampleResults/
└── {dataset}/
    ├── {sample_dir_name}/
    └── {dataset}_{sample_dir_name}.mat

• Each row in the mat contains the 5-dimensional quality scores for a distorted mesh sample.

---

3. Visualization

3.1 Visualize Mesh

python visualize_mesh.py \
    --dataset [tsmd|sjtumqa|yana] \
    --model_name vaso

Argument	Description
--dataset	Dataset name: tsmd, sjtumqa, or yana
--model_name	Model name to visualize

---

3.2 Visualize Local Patch SDF & NCF

python visualize_SDF_NCF.py \
    --dataset [tsmd|sjtumqa|yana] \
    --model_name vaso

Argument	Description
--dataset	Dataset name
--noise_level	Noise level: variance = (value × 0.01)²
--point_number	Number of points to visualize
--lp_number	Number of local patches
--model_name	Target model for visualization

---

4. Author & Contact

Kaifa Yang
Cooperative Medianet Innovation Center
Shanghai Jiao Tong University
✉️ Email: sekiroyyy@sjtu.edu.cn

---