Upgrade to Pro — share decks privately, control downloads, hide ads and more …

Video Propagation Networks

Video Propagation Networks

A presentation I gave about Video Propagation Networks at the 2018 edition of the "Advanced Methods in Computer Grpahics" Seminar at ETH Zurich.

See the original paper here: https://ps.is.tue.mpg.de/publications/vpn-cvpr17

Marcel Neidinger

March 23, 2018

More Decks by Marcel Neidinger

Other Decks in Research


  1. Beamerkalibrierungsuntersuchungstestvortragsfolie Quadrat Auflösung: 1920x1080 Stern in den Ecken zu sehen?

    Sowas wie 16:9? Nein? Soweit wie man es von Beamern erwarten kann? Quadrat quadratisch? Graustufen korrekt?
  2. „Talk is cheap - Show me the video“ - Linus

    Torvalds, Mastermind behind Linux
  3. Video Object Segmentation: Evaluation 0 20 40 60 80 SEA

    JMP BNN-Identity VPN Stage 2 VPN-Deeplab IoU F T 24
  4. Video Object Segmentation: Improvements 68 67,5 67 66,5 66 0

    500 1000 IoU Score Number of Points (in Thousand) 26
  5. Conclusion Three things to take away (i) VPNs use a

    bilateral and a spatial network (ii) Offer favorable runtime over current methods (iii) Perform better then current methods 1 1 And yes, this is the only textual slide you will see in this presentation
  6. Appendix • Definition of Performance Measures • Detailed Performance (Video

    Segmentation) • Detailed Performance (Semantic Labeling) • Detailed Performance (Color Propagation) • Literature Back to Overview • JumpCut (JMP) • SeamSeg (SEA)
  7. Performance Measures Intersection over Union (IoU) Groundtruth G, Output segmentation

    M Contour Accuracy F Temporal Stability T Back to Overview
  8. Literature (besides the main paper) [1] S. Paris and F.

    Durand. A fast approximation of the bilateralfilter using a signal processing approach. In European Conference on Computer Vision, pages 568–580. Springer, 2006. [2] V. Jampani, M. Kiefel, and P. V. Gehler. Learning sparse high dimensional filters: Image filtering, dense CRFs and bilateral neural networks. In Computer Vision and Pattern Recognition, IEEE Conference on, June 2016. Back to Overview
  9. JumpCut (JMP) Back to Overview • Key idea: Foreground and

    background exhibit different motions • Calculate two Nearest-Neighbor Fields to label target frame • Calculate Silhouette Edges (using the NNFs)
  10. SeamSeg (SEA) Back to Overview • Key idea: Seams(Connected Paths

    of low energy) Energy formulation allows for temporal label propagation