Deep Reading Notes: You Only Estimate Once

1. Research target

While conventional category-level pose often targets whole rigid objects, YOEO estimates functional parts within articulated objects. A cabinet may contain several drawer handles, requiring semantic segmentation, instance separation, and part pose together.

2. Three pointwise prediction tasks

For observed point $p_i$, predict:

• Semantic probability $s_i$: functional part class;
• Centroid offset $\Delta c_i$: direction to the instance center;
• NPCS coordinate $q_i$: location in the canonical part frame.

Voting points are:

$$\hat c_i=p_i+\Delta c_i.$$

Clustering same-semantic points in center space separates multiple handles of the same type. Then $(q_i,p_i)$ gives canonical-to-camera correspondences.

3. From NPCS to similarity transforms and part pose / size

Umeyama alignment fits a similarity transform:

$$\min_{s,R,t}\sum_i \left\|p_i-(sRq_i+t)\right\|_2^2.$$

Here $R$ is rotation, $t$ translation, and $s$ an isotropic scalar scale, rather than a 9D parameterization with independently regressed axis scales. RANSAC first rejects bad NPCS matches, preventing semantic-boundary errors from corrupting the closed-form fit.

The network predicts dense canonical correspondences rather than rotation directly; the geometric solver enforces a valid rigid / similarity transform.

4. Why parallel learning helps

Semantics, centroids, and NPCS are coupled:

• Semantics determine the canonical part frame;
• Centroids separate instances and reduce NPCS contamination from other same-class parts;
• NPCS supplies within-part geometry that can improve semantic boundaries.

Joint training lowers rotation error from 19.6 to 9.0 and raises mIoU from 52.3 to 57.6, supporting shared-representation synergy beyond speed alone.

5. Data and distribution

GAPartNet contains:

• 1166 objects;
• 8489 part instances;
• 27 object categories;
• 9 part classes;
• About 7.3 functional parts per object;
• About 8.8 object categories per part class.

Shared functional parts support cross-object-category generalization: drawer and door handles differ in appearance but share manipulation semantics and local geometry.

6. Results and robot validation

Real task	Successes / 10
Hinge handle	9
Drawer	5
Hinge lid	8
Total	22 / 30

GAPartNet achieves 20 / 30. Two additional successes suggest perception gains can aid manipulation, but 30 trials are few, and drawer success is only 50%; reliable real-world manipulation is not established.

7. Interpreting speed

Table I reports about 200 Hz from object-cloud input to part-pose output, covering the estimation path containing the first three stages below, without a per-stage timing breakdown. Raw RGB-D cloud generation, object detection / selection, and robot execution have a different timing scope. Distinguish:

• Backbone inference;
• point clustering;
• RANSAC + Umeyama;
• RGB-D cloud generation;
• Robot planning and control.

Object-cloud-to-part-pose inference at 200 Hz can support feedback, but does not mean the full raw-sensor-to-completed-grasp pipeline runs 200 times per second.

8. Limitations

• Few points on small parts destabilize centroid voting and NPCS.
• Metal / reflective surfaces cause missing RGB-D data.
• Canonical rotation of symmetric handles may be unobservable.
• Single-frame estimation omits joint kinematics and temporal continuity.
• Correct part pose does not guarantee collision-free reachability or a complete articulation trajectory.

9. Deep reading conclusion

YOEO concentrates learning into pointwise semantics and correspondences, with clustering and Umeyama enforcing instance / pose structure. Its strengths are part-level generalization and potential real-time feedback; robot experiments still reveal a substantial gap from 6D pose to reliable manipulation.