Summary
The paper studies scene flow estimation from point clouds, with emphasizing implicit rigidity constraints. Following the design of PointPWC-Net, the core of this approach lies in two primary components: a weight-sharing aggregation method and a deformation degree module. While the former leverages identical weights for the fusion of point coordinates, scene flow, and features to implicitly establish rigidity constraints, the latter emphasizes the preservation of local structures of rigid objects during the scene flow estimation process.
This paper showcases the efficacy of the resultant network, WSAFlowNet, by benchmarking its performance against the popular FlyingThings3D and KITTI datasets.
Strengths
- Overall, the writing of the manuscript is mostly clear and easy to follow. The authors provided various mathematical and experimental evidence to support the feasibility and effectiveness of their approach.
- In feature abstraction, the proposed weight-sharing aggregation module enforces rigidity constraints by using identical weights for aggregating point coordinates, scene flow, and features, without requiring explicit pose estimation or 3D object segmentation.
- The proposed deformation degree module reasonably enhances the local structure consistency by measuring the deformation degree of local structure between reference and warped points, with a KNN.
- Both proposed modules are properly ablated, which provides convincing signals with analysis.
Weaknesses
- Based on the ablations in Table 2, the gain from the deformation degree module looks minor. Indeed the two proposed modules work well in tandem, but more analysis and experimental evidence are needed to verify the necessity of the DD module for maintaining local rigid structure invariance.
- Relevant comparisons on the computation cost of different approaches could be useful for evaluating the practicality of the proposed model for the scene flow estimation task.
- No potential limitations or drawbacks of the proposed method are discussed. Also, failure cases need to be showcased and analyzed if any. Here the paper seems to focus too much on the advantages of the presented method and does not always give the whole picture.
- I'm not super familiar with recent SOTA methods on scene flow estimation from point clouds, so might be biased on this one: the generalization capability of the proposed model was demonstrated by evaluating on KITTI without finetuning, but it would be more convincing to demonstrate comparisons on more recent real-world point cloud data from autonomous driving datasets, such as nuScenes, Argoverse, and Waymo open dataset.
Questions
Please check the suggestions in the Weakness section above to further strengthen the paper.
Rating
6: marginally above the acceptance threshold
Confidence
3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.