Any ideas about the structure of unsupervised SUM-FCN

[chenbohua3]

After reading chapter 3.3 in FCSN several times, I can not figure out what exactly structure of the unsupervised part. Is that mean:

1. select Y frames: choose the top Y socres features with dimension:
   batch * 2 * Y
2. apply a 1*1 conv to decode features above to reconstruct their orginal feature representations:
   batch * 2 * Y -> batch * 10 * Y (shape of the output of conv8)
3. merge the input frame-level feature vectors of thess selected Y frames using skip connection:
   batch * 1024 * Y -> batch * 10 * Y
   and then added by the output of step 2
4. obtain final reconstructed features of the Y frames:
   batch * 10 * Y -> batch * 1024 * Y

[jiaweichuang]

Hello,have you implemented this unsupervised part?Can you tell me?