Easyportrait Eval

easyportrait-eval

EasyPortrait -- Face Parsing and Portrait Segmentation Dataset — Kapitanov et al. (2023) (arXiv:2304.13509, 2023)

What this evaluates

Evaluates semantic segmentation models on fine-grained face parsing and portrait segmentation. It probes a model's ability to accurately delineate nine distinct facial and occlusion classes in high-resolution indoor portrait images.

Datasets

• EasyPortrait — total 20000; splits: test (-1); repo https://github.com/hukenovs/easyportrait

Metrics

• mIoU (primary) — range: [0, 1]
  • Mean Intersection over Union computed across 9 segmentation classes. For each class, IoU is the intersection of predicted and ground truth masks divided by their union; mIoU is the average IoU across all classes.

Input / output format

Input: RGB portrait images resized to 1024 × 1024 pixels, padded to 1920 × 1920 during augmentation, with photometric distortions applied.

Output: 9-class semantic segmentation masks (background, person, face skin, eyes, eyebrows, lips, teeth, and occlusions like beard, glasses, hands).

Scoring recipe

def compute_miou(pred_masks, gt_masks, num_classes=9):
    ious = []
    for c in range(num_classes):
        pred_c = (pred_masks == c)
        gt_c = (gt_masks == c)
        intersection = np.logical_and(pred_c, gt_c).sum()
        union = np.logical_or(pred_c, gt_c).sum()
        if union == 0:
            ious.append(0.0)
        else:
            ious.append(intersection / union)
    return np.mean(ious)

Common pitfalls

• Applying random horizontal or vertical flips during training or evaluation, as it misaligns facial features like eyes and eyebrows.
• Using standard cross-entropy loss for BiSeNet-V2; the authors specify it requires OHEM loss.
• Training different architectures for different iteration counts (20,000 vs 40,000) and batch sizes (8 vs 16) to ensure fair comparison.

Evidence (verbatim from paper)

We report about it using mean Intersection over Union (mIoU). The results of our experiments (Table 4 and Table 5) show that all models trained on EasyPortrait have high enough values of the mIoU metric.

Citation

@misc{kapitanov2023easyportrait,
  title={EasyPortrait -- Face Parsing and Portrait Segmentation Dataset},
  author={Kapitanov et al. (2023)},
  year={2023},
  note={arXiv:2304.13509}
}

• arXiv: 2304.13509