What Matters for 3D Scene Flow Network

What Matters for 3D Scene Flow Network. Guangming Wang 1 , Yunzhe Hu 1 , Zhe Liu 1 , Yiyang Zhou 2 , Masayoshi Tomizuka 2 , Wei Zhan 2 , and Hesheng Wang 1 1 Shanghai Jiao Tong University 2 University of California, Berkeley.

Background. Task To predict point-wise displacement vectors from point clouds Motivation Previous flow-embedding based methods tend to miss correct yet distant matching point due to limited search range of K-NN. The estimated correspondence may not conform to the bidirectional consistency. The contributions to the performance gain of several designs of components and techniques are unclear for a typical scene flow network..

All-to-all flow embedding with backward validation.

All-to-all flow embedding with backward validation.

All-to-all flow embedding with backward validation.

All-to-all flow embedding with backward validation.

3D Scene Flow Network. 3 32 32 8 3 64 dlÖ•olll gaüMdji • 128 32 256 32 Flow Flow Refinement 512 Flow Refinement 128 Hierarchical Flow Refinement Flow Refinement Refinement Sja hxfixaotc; 32 256 128 3 3 128 3 256 64.

Alternatives of designs and techniques. Point similarity.

Experiments. Results on non-occluded dataset. Results on occluded dataset.

Experiments. Visualization. FlyingThings3D 1231 dataset KITTI Scene Flow 1251 dataset.