Memory-enhanced Categrory-Level Pose Estimation
Abstract
Category-level pose estimation jointly estimates the 6D pose: Rotation and translation, and object size for unseen objects with known category labels.
Currently, the SOTA methods in 9D are FS-Net [1] and
DualPoseNet? [2].
And one straightforward idea to improve the performance is to introduce priors into the network.
ShapePrior? [3] and CPS [4] leverage the point cloud to represent the mean shape of each category. FS-Net adopts the average size of each category.
We, instead, can use a memory module to store typical shapes of each category, similar to point cloud segmentation methods [5].
The way to establish the memory module:
1 First we train the network to extract features and then utilize the feature to reconstruct observed points, as in FS-Net.
2 Assume the features follow GM distribution, we can use a K-means to build the module, or some other unsupervised learning methods may be doable.
|
The way to train the network:
Our network structure is similar to FS-Net and
ShapePrior?, thus the training procedures may be similar too.
References:
[1] Chen, W., Jia, X., Chang, H. J., Duan, J., Shen, L., & Leonardis, A. (2021). FS-Net: Fast Shape-based Network for Category-Level 6D Object Pose Estimation with Decoupled Rotation Mechanism. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 1581-1590).
[2] Lin, J., Wei, Z., Li, Z., Xu, S., Jia, K., & Li, Y. (2021).
DualPoseNet?: Category-level 6D Object Pose and Size Estimation using Dual Pose Network with Refined Learning of Pose Consistency. arXiv preprint arXiv:2103.06526.
[3] Tian, M., Ang, M. H., & Lee, G. H. (2020, August). Shape Prior Deformation for Categorical 6D Object Pose and Size Estimation. In European Conference on Computer Vision (pp. 530-546). Springer, Cham.
[4] Manhardt, F., Wang, G., Busam, B., Nickel, M., Meier, S., Minciullo, L., ... & Navab, N. (2020). CPS++: Improving Class-level 6D Pose and Shape Estimation From Monocular Images With Self-Supervised Learning. arXiv preprint arXiv:2003.05848.
[5] He, T., Gong, D., Tian, Z., & Shen, C. (2020). Learning and Memorizing Representative Prototypes for 3D Point Cloud Semantic and Instance Segmentation. arXiv preprint arXiv:2001.01349
Resources
Literature
If you are interested in any of these topics, please contact us via e-mail.
Yan Di
Federico Tombari