Utd Mhad Eval

---
name: utd-mhad-eval
description: Evaluates a model's ability to predict future human joint positions over a 15-frame horizon using past observations, while testing continual learning capabilities across different subjects and curriculum-based fine-tuning. Use when the user wants to benchmark on UTD-MHAD, or asks about evaluating this task. Reports MSE.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2107.00544
  bibtex_key: yasar2021improving
  confidence: high
---

# utd-mhad-eval

> Improving Human Motion Prediction Through Continual Learning — Yasar et al. (2021) (arXiv:2107.00544, 2021)

## What this evaluates

Evaluates a model's ability to predict future human joint positions over a 15-frame horizon using past observations, while testing continual learning capabilities across different subjects and curriculum-based fine-tuning.

## Datasets

- **UTD-MHAD** — total ?; splits: train (-1), val (-1), test (-1)

## Metrics

- `MSE` **(primary)** — range: other
  - The $l_2$ distance between ground-truth and predicted poses at each timestep, averaged over the number of joints and sequence length. Formula: $\mathcal{L}(X,\hat{X}) = \frac{1}{T \cdot K} \sum_{t=1}^{T} \sum_{i=1}^{K} (x_t^i - \hat{x}_t^i)^2$, where T is total frames and K is total joints.

## Input / output format

**Input**: 15-frame sequence of 20-joint skeleton coordinates.

**Output**: 15-frame sequence of 20-joint skeleton coordinates.

## Scoring recipe

```python
def compute_mse(predictions, ground_truth):
    # predictions, ground_truth: shape (T=15, K=20)
    diff = predictions - ground_truth
    mse = np.mean(diff ** 2)
    return mse
```

## Common pitfalls

- The evaluation uses a strict cross-subject protocol (odd-numbered subjects for train/val, even-numbered for test) rather than random sequence splitting, which is essential for claiming generalization.
- Only skeleton data is used; RGB/depth modalities are explicitly excluded despite being available in the dataset.
- The zero-velocity baseline (predicting the last observed pose for all future frames) is a strong baseline for short-term prediction and must be reported alongside MSE to gauge movement dynamics.

## Evidence (verbatim from paper)

> We evaluated the performance of all models using the Mean Squared Error (MSE), which is the $l_{2}$ distance between the ground-truth and the predicted poses at each timestep, averaged over the number of joints and sequence length, in line with prior work. The MSE is calculated as: $\mathcal{L(X,\hat{X})}\=\frac{1}{T.K}\sum^{T}_{t\=1}\sum^{K}_{i\=1}(x_{t}^{i}-\hat{x}_{t}^{i})^{2}$ where, T and K are the total number of frame and joints respectively.

## Citation

```bibtex
@misc{yasar2021improving,
  title={Improving Human Motion Prediction Through Continual Learning},
  author={Yasar et al. (2021)},
  year={2021},
  note={arXiv:2107.00544}
}
```

- arXiv: 2107.00544