Dexycb Eval

dexycb-eval

DexYCB: A Benchmark for Capturing Hand Grasping of Objects — Chao et al. (2021) (arXiv:2104.04631, 2021)

What this evaluates

Evaluates joint perception and manipulation capabilities for hand-object interactions, specifically 2D detection, 6D object pose estimation, and 3D hand pose estimation on real-world RGB-D sequences.

Datasets

• DexYCB — total 582000; splits: train (-1), val (-1), test (-1)

Metrics

• precision-coverage (primary) — range: [0, 1]
  • Measures the trade-off between the precision of generated grasps and the coverage of successful grasp configurations across varying thresholds.
• 2D detection accuracy — range: [0, 1]
  • Standard object and keypoint detection metrics (e.g., AP) evaluated on RGB images.
• 6D object pose error — range: other
  • Translation and rotation error between predicted and ground-truth object poses.
• 3D hand pose error — range: other
  • Joint position error between predicted and ground-truth 3D hand poses.

Input / output format

Input: RGB-D frames captured from 8 synchronized camera views per frame.

Output: 2D bounding boxes and keypoints for hands/objects, 6D object pose matrices, 3D hand joint coordinates, and grasp proposals.

Scoring recipe

def compute_precision_coverage(pred_grasps, ref_grasps, threshold):
    correct = sum(1 for p in pred_grasps if any(distance(p, r) < threshold for r in ref_grasps))
    precision = correct / len(pred_grasps)
    coverage = correct / len(ref_grasps)
    return precision, coverage
# Plot precision vs coverage across thresholds

Common pitfalls

• Models trained on DexYCB's controlled lab background do not generalize to in-the-wild images (e.g., COCO), showing significant performance drops.
• Object-specific landmarks require explicit handling to scale labeling across diverse object geometries.
• Occlusions during hand-object interaction complicate 2D keypoint and 3D pose estimation.

Evidence (verbatim from paper)

Figure 9: Precision-coverage curves for grasp generation on S0, S2, and S3.

We tested a DexYCB-trained model on COCO images [21] and observed an expected drop in performance.

Citation

@misc{chao2021dexycb,
  title={DexYCB: A Benchmark for Capturing Hand Grasping of Objects},
  author={Chao et al. (2021)},
  year={2021},
  note={arXiv:2104.04631}
}

• arXiv: 2104.04631