Core Concepts
The author proposes STARFlow to address challenges in scene flow prediction by incorporating global attentive flow embedding and spatial temporal feature re-embedding, achieving state-of-the-art performance on various datasets.
Abstract
STARFlow introduces innovative modules to enhance scene flow prediction accuracy. The Global Attentive (GA) module matches point pairs globally, while the Spatial Temporal Feature Re-embedding (STR) module refines local features after deformation. Novel Domain Adaptive Losses bridge the gap between synthetic and real-world datasets, showcasing strong generalization. Experiments demonstrate superior performance across diverse datasets.
Key points:
Scene flow prediction is crucial for understanding dynamic scenes.
Challenges include local receptive fields and domain gaps.
STARFlow introduces GA for global matching and STR for local refinement.
Domain Adaptive Losses improve generalization to real-world datasets.
Achieves state-of-the-art performance on various datasets.
Stats
0.0143
0.0064
94ms / 9.88M
Quotes
"The proposed network leverages global attentive mechanisms."
"Our model achieves SOTA performance on multiple distinct datasets."