---
license: mit
tags:
  - federated-learning
  - knowledge-distillation
  - point-cloud
  - 3d-classification
  - benchmarking
  - privacy-preserving
  - distributed-training
datasets:
  - craniosynostosis
  - modelnet40
metrics:
  - accuracy
  - macro-accuracy
language:
  - en
model-index:
  - name: FLKD Benchmark Models
    results:
      - task:
          name: 3D Point Cloud Classification
          type: image-classification
        metrics:
          - name: Accuracy
            type: accuracy
          - name: Macro Accuracy
            type: accuracy
---

# Model Card: FLKD Point Cloud Classification Benchmark

## Model Details

### Model Family
PointNet++ (Single-Scale Grouping) with optional Federated Learning + Knowledge Distillation training

### Model Variants
- **Classical**: Standard centralized training
- **Federated**: Trained via 13 different federated learning algorithms (FedAvg, FedProx, SCAFFOLD, FedDyn, FedAvgM, FedAdam, FedYogi, FedAdagrad, FedMedian, FedBN, MOON, Ditto, FedNova)

### Input/Output

**Input**: 
- 3D point clouds with variable number of points (typically 1024 points)
- Float32 tensors of shape (batch_size, num_points, 3)

**Output**: 
- Class logits for binary classification (2 classes)
- Softmax probabilities for disease classification

## Intended Use

These models are designed for **3D point cloud classification**. They serve as benchmarks for evaluating:

1. **Federated Learning** effectiveness on 3D data with heterogeneous client distributions
2. **Knowledge Distillation** techniques applied to point cloud models
3. **Combined FL+KD** approaches for privacy-preserving distributed training
4. **Model performance variance** across multiple random seeds

### Primary Use Cases
- Research on federated learning for 3D deep learning
- Comparison of knowledge distillation losses on distributed systems
- Benchmark reference for new FL/KD algorithms on 3D data


## Model Performance


## Hardware & Training

### Training Environment
- **GPU**: NVIDIA GPUs; compatible with newer CUDA architectures
- **Distributed**: Single-GPU per job; parallel jobs via SLURM job arrays
- **Framework**: PyTorch 2.0+
## Recommendations

### How to Use
1. Load `best_model.pth` for production inference
2. Preprocess input point clouds to match training data distribution
3. Use model in evaluation mode (`model.eval()`) for inference
4. Report metrics across multiple seeds for reproducibility

### How to Cite
When using these models, please cite the original paper and dataset:

```bibtex
@inproceedings{aizierjiang26benchmark,
  title={Benchmarking Federated Learning and Knowledge Distillation for Point Cloud Classification},
  author={Aizierjiang Aiersilan},
  booktitle={European Conference on Computer Vision},
  organization={Springer},
  year={2026}
}
```

## Versioning

- **Model Version**: 1.0 (ECCV 2026 publication)
- **Framework**: PyTorch 2.0+
- **Python Version**: 3.10+

## Contact & Support

For questions about these models, please visit:
- **Project Website**: https://ezharjan.github.io/FLKD3DBenchmark/
- **GitHub Repository**: https://github.com/Ezharjan/FLKD3DBenchmark/

---

**Model Card Last Updated**: June 2026  
**Model Type**: Supervised Classification (3D Point Clouds)  
**Task**: Federated Learning + Knowledge Distillation Benchmark