Preview

1. 1e-4 - 1e-6 WarmupCosineAnnealing W/ Restart + Small DT (time issue)
2. context : recon = 1 : 2
3. context : recon = 1 : 6 (load ckpt 0008 of exp1)
4. context : recon = 1 : 6 + Only target region SSIM, LPIPS + GAN

Vision Model Enhancement (Fixed acoustic model)

loss 설명 (ratio)

context: Entire Depth (1) (exp1, 2 → 0.4)

recon: Target Depth (0.2)

perceptual: LPIPS (VGG16) Depth (exp1, 2 → Entire Depth) (0.1)

structural: SSIM Depth (exp1, 2 → Entire Depth) (0.1)

latent: MSE(latent, GT_depth) (0.005)

DepthGAN: PatchGAN Target Depth (0.01)

Summary