Stereo Ycb V Ds Eval

---
name: stereo-ycb-v-ds-eval
description: Evaluates 6D object pose estimation methods under stereo vision conditions, specifically probing robustness to occlusion and scale ambiguity by leveraging dense 2D-3D correspondences and stereo disparity. Use when the user wants to benchmark on Stereo PBR YCB-V DS, or asks about evaluating this task. Reports ADD0.1.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2402.05610
  bibtex_key: pollabauer2024extending
  confidence: high
---

# stereo-ycb-v-ds-eval

> Extending 6D Object Pose Estimators for Stereo Vision — Pöllabauer et al. (2024) (arXiv:2402.05610, 2024)

## What this evaluates

Evaluates 6D object pose estimation methods under stereo vision conditions, specifically probing robustness to occlusion and scale ambiguity by leveraging dense 2D-3D correspondences and stereo disparity.

## Datasets

- **Stereo PBR YCB-V DS** — total 481725; splits: train (433645), test (48080)

## Metrics

- `ADD0.1` **(primary)** — range: percent
  - Percentage of test frames where the average Euclidean distance between corresponding 3D points of the predicted and ground-truth poses is less than 10% of the object's diameter.

## Input / output format

**Input**: Stereo RGB image pairs (2048×1536) with corresponding depth maps (640×576), bounding boxes, and object masks for scenes containing up to 15 target objects.

**Output**: 6D pose (rotation and translation) for each target object in the frame.

## Scoring recipe

```python
def compute_add_at_0_1(pred_poses, gt_poses, diameters):
    correct = 0
    for pred, gt, d in zip(pred_poses, gt_poses, diameters):
        pred_pts = apply_pose(gt_pts, pred)
        gt_pts = apply_pose(gt_pts, gt)
        if np.mean(np.linalg.norm(pred_pts - gt_pts, axis=1)) < 0.1 * d:
            correct += 1
    return (correct / len(pred_poses)) * 100
```

## Common pitfalls

- Real-world test objects differ from synthetic training objects (e.g., texture/color changes), which can cause domain shift if not accounted for.
- Labels with less than 10% visible surface are removed, meaning the dataset is heavily filtered for stability and does not represent full occlusion scenarios.
- Monocular baselines must be extended to stereo (e.g., SO-Stereo, GDRN-Stereo) rather than simply fed stereo pairs, to ensure fair comparison.

## Evidence (verbatim from paper)

> Table 1: We report the performance of our proposed methods on a per-object basis, comparing the accuracy of a single multi-object model for both mono and stereo configurations using the common ADD0.1 metric.

## Citation

```bibtex
@misc{pollabauer2024extending,
  title={Extending 6D Object Pose Estimators for Stereo Vision},
  author={Pöllabauer et al. (2024)},
  year={2024},
  note={arXiv:2402.05610}
}
```

- arXiv: 2402.05610