A 3D CNN front-end plus temporal convolution and LSTM ensemble yields significant gains over baseline for active speaker detection on the AVA dataset.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.SD 1years
2019 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
Naver at ActivityNet Challenge 2019 -- Task B Active Speaker Detection (AVA)
A 3D CNN front-end plus temporal convolution and LSTM ensemble yields significant gains over baseline for active speaker detection on the AVA dataset.