实验

• 基本思路：整体级联、Stage1(E0+共享解码)、Stage2(E1+共享解码)

  I_{t-1} (HR 3x2H×2W)
     │  (运动对齐 / 外部warp)
     └─→ WarpHighColor_{t-1→t} (HR)
              │  S2D(scale=2) → 12ch
  i_t (LR 3) ─┐
  depth_t (1) ├──── concat → 16ch → Encoder E0 → Multi-scale feats → Shared Decoder → I_t_high (HR 3)
              │
              └──────────────────────────────────────────┐
  WarpLowColor_{t+1→t} (LR 3)                            |                   │                                       |
  I_t_high (HR 3) ── S2D → 12ch                          |                   │                                       |
  I_t_high bicubic → 3ch                                 |                   │                                       |
  depth_t (1)                                            |                   │                                       |
  concat → 19ch → Encoder E1 → feats → Shared Decoder → I_{t+1}_high (HR 3)
  

• E 0

  [ i_t | depth_t | S2D(WarpHighColor_{t-1→t}) ]  (B,16,H,W)
          │
     Stem: 3×3 Conv+PReLU → ResBlock → f1 (32 ch)
          │
          ├─ DownBlock(stride2) → 64 ch → f2
          │
          ├─ DownBlock(stride2) → 128 ch → f3
          │
          └─ DownBlock(stride2) → 256 ch → DepthGuidedAttention → f4
  
  Decoder(共享):
     f4 → Up(×2, ConvT) → 128  ─ Add f3 → Refine(Conv+Res) → 128a
     128a → Up → 64 ─ Add f2 → Refine → 64a
     64a → Up → 32 ─ Add f1 → Refine → 32a
     32a ─ Concat[ + f1 ] → Conv → 46 → Conv → 12 → PixelShuffle(scale=2) → I_t_high (B,3,2H,2W)
  收集解码特征: [128a,64a,32a]
  

• E 1