แนวคิดหลัก
3D hand pose estimation benefits from single-to-dual-view adaptation for improved accuracy and flexibility.
บทคัดย่อ
The pursuit of accurate 3D hand pose estimation in egocentric vision drives the need for adapting single-view estimators to dual views. The proposed Single-to-Dual-View adaptation (S2DHand) method eliminates the need for multi-view annotations during training and can handle arbitrary dual-view pairs with unknown camera parameters. By leveraging stereo constraints, the model achieves significant improvements in both in-dataset and cross-dataset settings, outperforming existing methods.
- Introduction:
- Accurate 3D hand pose estimation is crucial for understanding human activity in egocentric vision.
- Existing methods rely on single-view images, leading to limitations in field-of-view and depth ambiguity.
- Problem Setting:
- The task involves adapting a pre-trained single-view hand pose estimator to arbitrary dual views without multi-view labels or camera parameters.
- Proposed Method:
- S2DHand utilizes stereo constraints for adaptation, including cross-view consensus and invariance of transformation.
- Experiment:
- Evaluation on dual-camera pairs shows significant improvements in accuracy under both in-dataset and cross-dataset settings.
- Related Work:
- Comparison with state-of-the-art adaptation methods demonstrates the superiority of S2DHand in cross-dataset scenarios.
- Ablation Study:
- Components like attention-based merging and rotation-guided refinement contribute to the overall performance of the method.
- Hyper-parameters:
- Optimal values for α and β are crucial for achieving the best performance in the adaptation process.
- Qualitative Result:
- Visual examples showcase the improved accuracy of 3D hand pose estimation under dual-view settings.
สถิติ
"Our method achieves the best performance when α = 0.7."
"S2DHand achieves the best performance when β = ∞."
คำพูด
"Our method achieves significant improvements on arbitrary camera pairs under both in-dataset and cross-dataset settings."