CD Fer Benchmark Eval

cd-fer-benchmark-eval

Cross-Domain Facial Expression Recognition: A Unified Evaluation Benchmark and Adversarial Graph Learning — Chen et al. (2020) (arXiv:2008.00923, 2020)

What this evaluates

Evaluates cross-domain facial expression recognition (CD-FER) models by measuring how well they transfer learned features from a labeled source dataset to an unlabeled target dataset. It probes the model's ability to learn domain-invariant representations and adapt to distribution shifts across different facial expression datasets.

Datasets

• RAF-DB — total ?; splits: train (-1), test (-1)
• AFE — total ?; splits: train (-1), test (-1)
• CK+ — total ?; splits: train (-1), test (-1)
• JAFFE — total ?; splits: train (-1), test (-1)
• SFEW2.0 — total ?; splits: train (-1), test (-1)
• FER2013 — total ?; splits: train (-1), test (-1)
• ExpW — total ?; splits: train (-1), test (-1)

Metrics

• accuracy (primary) — range: percent
  • Percentage of correctly predicted facial expression labels out of the total number of samples in the target domain.

Input / output format

Input: Facial images from a labeled source domain and an unlabeled target domain, processed through a backbone network (e.g., ResNet-50, ResNet-18, MobileNet-v2) to extract holistic and local features.

Output: Predicted facial expression labels for target domain images. Evaluation reports per-dataset accuracy and mean accuracy across all target datasets.

Scoring recipe

correct = 0
total = 0
for img, label in target_dataset:
    pred = model(img)
    if pred == label:
        correct += 1
    total += 1
accuracy = (correct / total) * 100
mean_accuracy = sum(accuracies_per_dataset) / num_datasets

Common pitfalls

• Using inconsistent source/target dataset or backbone combinations without averaging fairly across methods can lead to biased performance comparisons.
• Relying solely on holistic features without local feature co-adaptation significantly degrades cross-domain transfer performance, as holistic features alone lack domain-invariant cues.
• Failing to update per-class statistical distributions or fix adjacency matrices during training causes severe accuracy drops due to poor node initialization and message propagation.

Evidence (verbatim from paper)

Specifically, when using the RAF source dataset and ResNet-18 backbone, our AGRA approach obtains accuracies of 77.52%, 61.03%, 52.75%, 54.94%, 69.70% on the CK+, JAFFE, SFEW2.0, FER2013, ExpW datasets, outperforming all of the current best-performing methods.

Citation

@misc{chen2020cdfer,
  title={Cross-Domain Facial Expression Recognition: A Unified Evaluation Benchmark and Adversarial Graph Learning},
  author={Chen et al. (2020)},
  year={2020},
  note={arXiv:2008.00923}
}

• arXiv: 2008.00923