zhenglinpan / FI3D

Revisiting Frechet Inception Distance for 3D Human Motion Generation
MIT License
0 stars 0 forks source link

Discussion with Gohar #20

Closed zhenglinpan closed 1 year ago

zhenglinpan commented 1 year ago

Take-aways:

  1. Re-definition of the biased test: change Encoder instead of the main branch
  2. VAE: to regulate the distribution of the Encoder, make reliable of FID
zhenglinpan commented 12 months ago

Further discussion with Ali, We can proceed with two approaches:

Encoder(instrinsic approach)

Use some techniques that allows current encoder to generate naturally normal distributed feature maps. I.e. VAE

Data(extrinsic approach)

Use some techiniques on generated data distribution to make it normal distributed in a way. i.e. Expectation Maximization, or Decomposition.