Densepose Coco Eval

densepose-coco-eval

DensePose: Dense Human Pose Estimation In The Wild — Güler et al. (2018) (arXiv:1802.00434, 2018)

What this evaluates

Evaluates a model's ability to perform dense human pose estimation by predicting per-pixel body part labels and UV coordinates on a 3D surface model. It measures how well the model handles real-world variations in scale, pose, occlusion, and background clutter.

Datasets

• COCO-DensePose — total ?; splits: train (48000), test (1500)

Metrics

• AP (primary) — range: [0, 100]
  • Average Precision computed over predicted instances and ground truth, averaged across IoU thresholds (0.50:0.05:0.95) and object sizes (small, medium, large).

Input / output format

Input: RGB images containing humans. Single-person evaluation uses crops around ground-truth boxes; multi-person uses full in-the-wild images.

Output: Per-instance bounding boxes, body part labels, and UV coordinates mapped to a 3D parametric surface (SMPL).

Scoring recipe

def compute_ap(preds, gold, iou_thresh=0.5):
    tp, fp = 0, 0
    for p in sorted(preds, key=lambda x: x['score'], reverse=True):
        if any(iou(p['box'], g['box']) >= iou_thresh for g in gold):
            tp += 1
        else:
            fp += 1
    return tp / (tp + fp) if (tp + fp) > 0 else 0.0
# AP is averaged over IoU thresholds 0.50:0.05:0.95 and object sizes (S, M, L)

Common pitfalls

• Using ground-truth human masks to remove background during evaluation inflates performance but breaks comparability with standard in-the-wild settings.
• Multi-scale testing and ensembling significantly boost scores but are not part of the standard single-scale evaluation protocol.

Evidence (verbatim from paper)

In Table 1 we report the AP and AR metrics described in Sec. 2 as we change different choices in our architecture.

Citation

@misc{guler2018densepose,
  title={DensePose: Dense Human Pose Estimation In The Wild},
  author={Güler et al. (2018)},
  year={2018},
  note={arXiv:1802.00434}
}

• arXiv: 1802.00434