Algorithm 1: Enhanced Multi-Modal Feature Fusion at scale  $s$ (Equations (2)–(15))

Input:  RGB image $I\in R^{H\times W\times 3}$  Sparse depth map $D\in R^{H\times W\times 1}$  Pre-processed dense depth (from BP) $D_s^{\left(pre\right)}\in R^{H\times W}$  Camera intrinsics $K\in R^{3\times 3}$ Output:  Fused feature map $F_s\in R^{H\times W\times C}$ Notion:  $\oplus$: channel concatenation;  $\odot$: channel-wise multiplication;  $GAP$: global average pooling;  $\delta (\cdot )$: ReLU;  $\sigma (\cdot )$: Sigmoid  $D_s^{\left(pre\right)}$: dense depth pre−processed via Bilateral Propagation Procedure: 1: Step 1: Back-projection to 3D 2:  Construct sparse point cloud $P=\mathrm{B}\mathrm{a}\mathrm{c}\mathrm{k}\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{j}\mathrm{e}\mathrm{c}\mathrm{t}\left(S,K\right)$. 3: Step 2: 2D feature encoding 4:  Form 2D input $X=[I\oplus D_s^{\left(pre\right)}]$ 5:  $F_{2D}=U-Net\left(X\right)$ 6: Step 3: Dual-branch PointNet++-S 7:  For each branch $b\in \left\{local,context\right\}$: 8:   For each SA layer with config $(n_l,r_l,k_l)$: 9:    $g_l^{\left(b\right)}\leftarrow \mathrm{PointNetSA}(P;n_l,r_l,k_l)$ 10:   End for 11:  End for 12:  Aggregate to multi-scale 3D feature $F_{3D}=\mathrm{Concat}\left(F_{\mathrm{local}}^{\left(2\right)},F_{\mathrm{context}}^{\left(2\right)}\right)$ 13: Step 4: Channel recalibration on 3D 14:  $z=\mathrm{GAP}\left(F_{3D}\right)$; 15:  $s=\sigma \left(W_2\cdot \delta \left(W_{1}z\right)\right)$; 16:  $G_{3D}=s\odot F_{3D}$. 17: Step 5: Spatial broadcasting of 3D feature 18:  $\widetilde{G_{3D}}=\mathrm{Broadcast}\left(G_{3D}\to H\times W\right)$ 19: Step 6: CAFFM: channel-attention fusion 20:  $F_{\mathrm{in}}=\mathrm{Concat}\left(F_{2D},\widetilde{G_{3D}}\right)$; 21:  $z=\mathrm{GAP}\left(F_{\mathrm{in}}\right)$; 22:  $\alpha =\sigma \left(W_2\cdot \delta \left(W_{1}z\right)\right)$); 23:  $F_s=\alpha \odot F_{\mathrm{in}}$. 24: return $F_s$.