FlowIt: Global Matching for Optical Flow with Confidence-Guided Refinement

Dimensions overall score 7.0

GitHub Code Pulse

No public code linked for this paper yet.

• FlowIt achieves state-of-the-art results on the Sintel benchmark.result 90%
• FlowIt achieves state-of-the-art results on the KITTI benchmark.result 90%
• FlowIt establishes new state-of-the-art cross-dataset zero-shot generalization performance on Sintel, Spring, and LayeredFlow.result 90%
• FlowIt formulates flow initialization as an optimal transport problem to produce a robust initial flow field, occlusion map, and confidence map.method 95%
• FlowIt uses a confidence-guided refinement stage that propagates reliable motion estimates from high-confidence regions into low-confidence areas.method 90%
• FlowIt leverages a hierarchical transformer architecture to capture extensive global context for handling large pixel displacements.technical 90%
• The initial confidence map is derived from the optimal transport probability map by summing matching probabilities within a local window surrounding the peak.technical 95%
• FlowIt (XL) outperforms the best WAFT variant (DAv2-a2) on the KITTI benchmark, reducing Fl-all error from 2.36/1.36/2.75 to 1.62/0.76/2.47 across metrics.result 95%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card