Use AbsGrad to get better results with less gaussians (#3113)

* Use AbsGrad to get better results with less gaussians

* some adjustment such that splatfacto-big don't go OOM

* use new gsplat version

---------

Co-authored-by: Jianbo Ye <jianboye@amazon.com>
Co-authored-by: Justin Kerr <justin.g.kerr@gmail.com>

nerfstudio/configs/method_configs.py

@@ -658,6 +658,7 @@ method_configs["splatfacto-big"] = TrainerConfig(
         model=SplatfactoModelConfig(
             cull_alpha_thresh=0.005,
             continue_cull_post_densification=False,
+            densify_grad_thresh=0.0006,
         ),
     ),
     optimizers={

nerfstudio/models/splatfacto.py

@@ -80,6 +80,21 @@ def SH2RGB(sh):
     return sh * C0 + 0.5
 
 
+def resize_image(image: torch.Tensor, d: int):
+    """
+    Downscale images using the same 'area' method in opencv
+
+    :param image shape [H, W, C]
+    :param d downscale factor (must be 2, 4, 8, etc.)
+
+    return downscaled image in shape [H//d, W//d, C]
+    """
+    import torch.nn.functional as tf
+
+    weight = (1.0 / (d * d)) * torch.ones((1, 1, d, d), dtype=torch.float32, device=image.device)
+    return tf.conv2d(image.permute(2, 0, 1)[:, None, ...], weight, stride=d).squeeze(1).permute(1, 2, 0)
+
+
 @dataclass
 class SplatfactoModelConfig(ModelConfig):
     """Splatfacto Model Config, nerfstudio's implementation of Gaussian Splatting"""
@@ -103,7 +118,7 @@ class SplatfactoModelConfig(ModelConfig):
     """If True, continue to cull gaussians post refinement"""
     reset_alpha_every: int = 30
     """Every this many refinement steps, reset the alpha"""
-    densify_grad_thresh: float = 0.0002
+    densify_grad_thresh: float = 0.0008
     """threshold of positional gradient norm for densifying gaussians"""
     densify_size_thresh: float = 0.01
     """below this size, gaussians are *duplicated*, otherwise split"""
@@ -379,8 +394,8 @@ class SplatfactoModel(Model):
         with torch.no_grad():
             # keep track of a moving average of grad norms
             visible_mask = (self.radii > 0).flatten()
-            assert self.xys.grad is not None
-            grads = self.xys.grad.detach().norm(dim=-1)
+            assert self.xys.absgrad is not None  # type: ignore
+            grads = self.xys.absgrad.detach().norm(dim=-1)  # type: ignore
             # print(f"grad norm min {grads.min().item()} max {grads.max().item()} mean {grads.mean().item()} size {grads.shape}")
             if self.xys_grad_norm is None:
                 self.xys_grad_norm = grads
@@ -631,12 +646,7 @@ class SplatfactoModel(Model):
     def _downscale_if_required(self, image):
         d = self._get_downscale_factor()
         if d > 1:
-            newsize = [image.shape[0] // d, image.shape[1] // d]
-
-            # torchvision can be slow to import, so we do it lazily.
-            import torchvision.transforms.functional as TF
-
-            return TF.resize(image.permute(2, 0, 1), newsize, antialias=None).permute(1, 2, 0)
+            return resize_image(image, d)
         return image
 
     def get_outputs(self, camera: Cameras) -> Dict[str, Union[torch.Tensor, List]]:
@@ -746,15 +756,10 @@ class SplatfactoModel(Model):
 
             return {"rgb": rgb, "depth": depth, "accumulation": accumulation, "background": background}
 
-        # Important to allow xys grads to populate properly
-        if self.training:
-            self.xys.retain_grad()
-
         if self.config.sh_degree > 0:
             viewdirs = means_crop.detach() - optimized_camera_to_world.detach()[:3, 3]  # (N, 3)
-            viewdirs = viewdirs / viewdirs.norm(dim=-1, keepdim=True)
             n = min(self.step // self.config.sh_degree_interval, self.config.sh_degree)
-            rgbs = spherical_harmonics(n, viewdirs, colors_crop)
+            rgbs = spherical_harmonics(n, viewdirs, colors_crop)  # input unnormalized viewdirs
             rgbs = torch.clamp(rgbs + 0.5, min=0.0)  # type: ignore
         else:
             rgbs = torch.sigmoid(colors_crop[:, 0, :])
@@ -762,7 +767,6 @@ class SplatfactoModel(Model):
         assert (num_tiles_hit > 0).any()  # type: ignore
 
         # apply the compensation of screen space blurring to gaussians
-        opacities = None
         if self.config.rasterize_mode == "antialiased":
             opacities = torch.sigmoid(opacities_crop) * comp[:, None]
         elif self.config.rasterize_mode == "classic":
@@ -923,11 +927,7 @@ class SplatfactoModel(Model):
         gt_rgb = self.composite_with_background(self.get_gt_img(batch["image"]), outputs["background"])
         d = self._get_downscale_factor()
         if d > 1:
-            # torchvision can be slow to import, so we do it lazily.
-            import torchvision.transforms.functional as TF
-
-            newsize = [batch["image"].shape[0] // d, batch["image"].shape[1] // d]
-            predicted_rgb = TF.resize(outputs["rgb"].permute(2, 0, 1), newsize, antialias=None).permute(1, 2, 0)
+            predicted_rgb = resize_image(outputs["rgb"], d)
         else:
             predicted_rgb = outputs["rgb"]
 

pyproject.toml

@@ -63,7 +63,7 @@ dependencies = [
     "xatlas",
     "trimesh>=3.20.2",
     "timm==0.6.7",
-    "gsplat>=0.1.9",
+    "gsplat>=0.1.11",
     "pytorch-msssim",
     "pathos",
     "packaging"