Facial Attribute Prediction Eval

facial-attribute-prediction-eval

15M Multimodal Facial Image-Text Dataset — Dawei Dai et al. (2024) (arXiv:2407.08515, 2024)

What this evaluates

Tests the model's capability to predict multiple facial attributes (e.g., gender, hairstyle) from a single facial image as a multilabel classification task. It probes fine-grained visual feature extraction and attribute-level alignment.

Datasets

• CelebA — total 182732; splits: train (162770), test (19962)
• LFWA — total ?; splits: train (6263), test (-1)

Metrics

• Average Precision (AP) (primary) — range: percent
  • Mean Average Precision across all facial attributes. Computes the area under the precision-recall curve for each label and averages them across the attribute set.

Input / output format

Input: A single facial image.

Output: A set of predicted binary labels for each facial attribute.

Scoring recipe

def compute_ap(predictions, gold):
    aps = []
    for pred_labels, true_labels in zip(predictions, gold):
        tp = 0
        precisions, recalls = [], []
        for i, pred in enumerate(pred_labels):
            if pred == true_labels[i]: tp += 1
            precisions.append(tp / (i + 1))
            recalls.append(tp / sum(true_labels))
        aps.append(np.trapz(precisions, recalls))
    return np.mean(aps) * 100

Common pitfalls

• Evaluates across different training data scales (1%, 2%/10%, 100%), so results are not directly comparable without noting the subset size.
• Uses multiple feature aggregation strategies (CLS, mean, max pooling) combined via layer normalization before the final linear layer, which can inflate performance if not standardized.

Evidence (verbatim from paper)

Average precision (AP) was used as the evaluation index.

Citation

@misc{dai2024facecaption,
  title={15M Multimodal Facial Image-Text Dataset},
  author={Dawei Dai et al. (2024)},
  year={2024},
  note={arXiv:2407.08515}
}

• arXiv: 2407.08515