Fork me on GitHub

OSVOS-S: Video Object Segmentation
Without Temporal Information

Publication

K.K. Maninis*, S. Caelles*, Y.Chen, J. Pont-Tuset, L. Leal-Taixé, D. Cremers, and L. Van Gool
Video Object Segmentation Without Temporal Information, Transactions of Pattern Analysis and Machine Intelligence (T-PAMI), 2018.
[BibTeX] [PDF] 
@Article{Man+18b,
  Author 	= {Kevis-Kokitsi Maninis and Sergi Caelles and Yuhua Chen and Jordi Pont-Tuset and Laura Leal-Taix\'e and Daniel Cremers and Luc {Van Gool}},
  Title 	= {Video Object Segmentation Without Temporal Information},
  Journal 	= {IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
  Year 		= {2018}
}

Team

Method

We improve on OSVOS by plugging in instance-aware semantic information, coming from an instance segmentation method (MNC, FCIS, MaskRCNN). The overall pipeline is illustrated in the figure below:

Results

The table shows the overall results of OSVOS-S compared to the state of the art in the validation set of DAVIS. Below the per-sequence results of OSVOS-S compared to the previous state of the art.
In terms of speed, below the plot of quality versus time per frame.
We reach the same quality as OFL in 160 miliseconds vs 40 seconds,
and the same quality as OnAVOS in 1.1 seconds vs 10 seconds.

For a complete comparison and qualitative results, visit the DAVIS website.

Downloads