RVOS: Proposal-generation, Refinement and Merging for Video Object Segmentation

Carles Ventura et al., CVPR, 2019 [paper] [project page] [poster]


191021-rvos-fig1

Approach

191021-rvos-fig2

1. Encoder

2. Decoder

191021-rvos-fig3

where $B_2$ is the bilnear upsampling operater by a factor 2,
$f^\prime_{t,k}$ is the result of projecting $f_{t,k}$ to have low dimensionality via conv. layer.
equation \eqref{eq:3}은 $k\in{1,\dots,n_b}$(number of conv. block)에 대해 chain으로 적용된다.

한편 $h_{t,i,0}$과 $h_{state}$는 다음과 같이 얻어진다.

where $Z$ = zero matrix. (= no previous spatial hidden state)

다시한번 그림을 보면서 식을 음미해 봅시다.

191021-rvos-fig3

Experiments and result

Spatial Recurrence vs. Spatio-temporal recurrence

191022-rvos-tab1

Overall Results

191022-rvos-tab2

191022-rvos-tab3

Zeroshot Results

191022-rvos-tab5

장점:
한계:

Comments

Eungbean Lee's Picture

About Eungbean Lee

Lee is a Student, Programmer, Engineer, Designer and a DJ

Seoul, South Korea https://eungbean.github.io