remove redundant code

README.md

@@ -152,7 +152,7 @@ We thank the kindly help of [David Charatan](https://davidcharatan.com/#/) for p
 @article{ye2024noposplat,
       title   = {No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Images},
       author  = {Ye, Botao and Liu, Sifei and Xu, Haofei and Xueting, Li and Pollefeys, Marc and Yang, Ming-Hsuan and Songyou, Peng},
-      journal = {arXiv preprint arXiv:xxxx.xxxx},
+      journal = {arXiv preprint arXiv:2410.24207},
       year    = {2024}
     }
 ```

src/model/decoder/cuda_splatting.py

@@ -11,7 +11,6 @@
 from torch import Tensor
 
 from ...geometry.projection import get_fov, homogenize_points
-from ..encoder.epipolar.conversions import depth_to_relative_disparity
 
 
 def get_projection_matrix(
@@ -229,49 +228,3 @@ def render_cuda_orthographic(
 
 
 DepthRenderingMode = Literal["depth", "disparity", "relative_disparity", "log"]
-
-
-def render_depth_cuda(
-    extrinsics: Float[Tensor, "batch 4 4"],
-    intrinsics: Float[Tensor, "batch 3 3"],
-    near: Float[Tensor, " batch"],
-    far: Float[Tensor, " batch"],
-    image_shape: tuple[int, int],
-    gaussian_means: Float[Tensor, "batch gaussian 3"],
-    gaussian_covariances: Float[Tensor, "batch gaussian 3 3"],
-    gaussian_opacities: Float[Tensor, "batch gaussian"],
-    scale_invariant: bool = True,
-    mode: DepthRenderingMode = "depth",
-) -> Float[Tensor, "batch height width"]:
-    # Specify colors according to Gaussian depths.
-    camera_space_gaussians = einsum(
-        extrinsics.inverse(), homogenize_points(gaussian_means), "b i j, b g j -> b g i"
-    )
-    fake_color = camera_space_gaussians[..., 2]
-
-    if mode == "disparity":
-        fake_color = 1 / fake_color
-    elif mode == "relative_disparity":
-        fake_color = depth_to_relative_disparity(
-            fake_color, near[:, None], far[:, None]
-        )
-    elif mode == "log":
-        fake_color = fake_color.minimum(near[:, None]).maximum(far[:, None]).log()
-
-    # Render using depth as color.
-    b, _ = fake_color.shape
-    result = render_cuda(
-        extrinsics,
-        intrinsics,
-        near,
-        far,
-        image_shape,
-        torch.zeros((b, 3), dtype=fake_color.dtype, device=fake_color.device),
-        gaussian_means,
-        gaussian_covariances,
-        repeat(fake_color, "b g -> b g c ()", c=3),
-        gaussian_opacities,
-        scale_invariant=scale_invariant,
-        use_sh=False,
-    )
-    return result.mean(dim=1)

src/model/encoder/encoder_noposplat_multi.py

@@ -19,8 +19,6 @@
 from .backbone import Backbone, BackboneCfg, get_backbone
 from .common.gaussian_adapter import GaussianAdapter, GaussianAdapterCfg, UnifiedGaussianAdapter
 from .encoder import Encoder
-from .epipolar.depth_predictor_monocular import DepthPredictorMonocular
-from .epipolar.epipolar_transformer import EpipolarTransformer, EpipolarTransformerCfg
 from .visualization.encoder_visualizer_epipolar_cfg import EncoderVisualizerEpipolarCfg
 
 
@@ -63,10 +61,7 @@ def rearrange_head(feat, patch_size, H, W):
 
 class EncoderNoPoSplatMulti(Encoder[EncoderNoPoSplatCfg]):
     backbone: nn.Module
-    depth_predictor: DepthPredictorMonocular
-    to_gaussians: nn.Sequential
     gaussian_adapter: GaussianAdapter
-    high_resolution_skip: nn.Sequential
 
     def __init__(self, cfg: EncoderNoPoSplatCfg) -> None:
         super().__init__(cfg)
@@ -167,22 +162,16 @@ def forward(
                 all_mean_res.append(res2)
 
             # for the 3DGS heads
-            if self.gs_params_head_type == 'linear':
-                GS_res1 = rearrange_head(self.gaussian_param_head(dec1[-1]), self.patch_size, h, w)
-                GS_res2 = rearrange_head(self.gaussian_param_head2(dec2[-1]), self.patch_size, h, w)
-            elif self.gs_params_head_type == 'dpt':
-                GS_res1 = self.gaussian_param_head([tok.float() for tok in dec1], shape1[0].cpu().tolist())
-                GS_res1 = rearrange(GS_res1, "b d h w -> b (h w) d")
-                GS_res2 = self.gaussian_param_head2([tok.float() for tok in dec2], shape2[0].cpu().tolist())
-                GS_res2 = rearrange(GS_res2, "b d h w -> b (h w) d")
-            elif self.gs_params_head_type == 'dpt_gs':
+            if self.gs_params_head_type == 'dpt_gs':
                 GS_res1 = self.gaussian_param_head([tok[:, 0].float() for tok in dec_feat], all_mean_res[0]['pts3d'].permute(0, 3, 1, 2), images[:, 0, :3], shape[0, 0].cpu().tolist())
                 GS_res1 = rearrange(GS_res1, "b d h w -> b (h w) d")
                 all_other_params.append(GS_res1)
                 for i in range(1, v):
                     GS_res2 = self.gaussian_param_head2([tok[:, i].float() for tok in dec_feat], all_mean_res[i]['pts3d'].permute(0, 3, 1, 2), images[:, i, :3], shape[0, i].cpu().tolist())
                     GS_res2 = rearrange(GS_res2, "b d h w -> b (h w) d")
                     all_other_params.append(GS_res2)
+            else:
+                raise NotImplementedError(f"unexpected {self.gs_params_head_type=}")
 
         pts_all = [all_mean_res_i['pts3d'] for all_mean_res_i in all_mean_res]
         pts_all = torch.stack(pts_all, dim=1)