Embedded Files
Denis Hadjivelichkov, Sicelukwanda Zwane, Marc Peter Deisenroth, Lourdes Agapito, Dimitrios Kanoulas
Current pose correspondence methods rely on vast 3D object datasets and do not actively consider semantic information, which limits the objects to which they can be applied. We present a novel method for learning cross-object pose correspondence. The proposed method detects interacting object parts, performs one-shot part correspondence, and uses geometric and visual-semantic features. Given one example of two objects posed relative to each other, the model can learn how to transfer the demonstrated relations to unseen object instances.
Supplementary Results
Supplementary Results
Correspondence Quality
We compare the output quality of our model and the baseline model. Each row represents the mug used in the demonstrated example, and each coolumn represents the mug used for imitation. Diagonal cells depict imitation with the same tool as used in the example interaction.
Our model
TAX-Pose
Survey.
We conducted a survey assessing the quality of the predictions of our method and TAX-Pose baseline. All surveyed users have some technical knowledge, such as academics, analysts, and engineers. Their expertise ensures they can provide informed evaluations of the methods and predictions being compared without specific knowledge of the subject matter. None of the participants were directly involved in the development of the methods under comparison. This ensures that their assessments are unbiased and objective, providing a fair evaluation of the methods' effectiveness. The participants' diverse perspectives contribute to a comprehensive evaluation of the methods.
Evaluation was done using a standard scale of 1-5 with extremes:
1 is “does not resemble the demonstrated interaction at all”
5 is “it resembles the demonstrated interaction as closely as possible”
And the interpretation of this was left to the individual surveyed person.
Each survey includes 18 responses (9 for each model). In total 24 people responded to the survey.
Mug on Rack. Following are examples of our model's mug-on rack one-shot predictions
Example Failure Modes
Failed prediction is attributed to the small part on which the tool is acting
Failed prediction in mug-pour - a small gap between the mugs causes 0% success rate
The image depicts a demonstrated example interaction of a plastic hammer and a nail on a bench. Imperfect pointclouds make one-shot use of our model in reality difficult.
Authors
Authors
Page updated
Google Sites
Report abuse