Speedup by merging repeated computation

README.md

@@ -21,9 +21,12 @@
 </p>
 
 <p align="center">
-Gaussian Opacity Fields (GOF) enables geometry extraction with 3D Gaussians directly by indentifying its level set. Our regularization improves surface reconstruction and we utilize Marching Tetrahedra for compact and adaptive mesh extraction.</p>
+Gaussian Opacity Fields (GOF) enables geometry extraction with 3D Gaussians directly by indentifying its level set. Our regularization improves surface reconstruction and we utilize Marching Tetrahedra for adaptive and compact mesh extraction.</p>
 <br>
 
+# Updates
+
+* **[2024.06.10]**: 🔥 Improve the training speed by 2x with [merged operations](https://github.com/autonomousvision/gaussian-opacity-fields/pull/58). 6 scenes in TNT dataset can be trained in ~24 mins and the bicycle scene in the Mip-NeRF 360 dataset can be trained in ~45 mins. Please pull the latest code and reinstall with `pip install submodules/diff-gaussian-rasterization` to use it.
 
 # Installation
 Clone the repository and create an anaconda environment using

evaluate_dtu_mesh.py

@@ -130,13 +130,13 @@ def cull_mesh(cameras, mesh):
     mesh.update_faces(face_mask)
 
     # Taking the biggest connected component
-    print("Taking the biggest connected component")
-    components = mesh.split(only_watertight=False)
-    areas = np.array([c.area for c in components], dtype=np.float32)
-    mesh_clean = components[areas.argmax()]
-
-    return mesh_clean
+    # print("Taking the biggest connected component")
+    # components = mesh.split(only_watertight=False)
+    # areas = np.array([c.area for c in components], dtype=np.float32)
+    # mesh_clean = components[areas.argmax()]
 
+    # return mesh_clean
+    return mesh
 
 def evaluate_mesh(dataset : ModelParams, iteration : int, DTU_PATH : str):
 

scene/gaussian_model.py

@@ -196,7 +196,20 @@ def get_view2gaussian(self, viewmatrix):
 
         # transpose view2gaussian to match glm in CUDA code
         V2G = V2G.transpose(2, 1).contiguous()
-        return V2G
+
+        # precompute results to reduce computation and IO
+        scales = self.get_scaling_with_3D_filter
+        S_inv_square = 1.0 / (scales ** 2)
+        R = V2G[:, :3, :3].transpose(1, 2)
+        t2 = V2G[:, 3:, :3]
+
+        C = torch.sum((t2 ** 2) * S_inv_square[:, None, :], dim=2)
+        S_inv_square_R = S_inv_square[:, :, None] * R
+        B = t2 @ S_inv_square_R
+        Sigma = R.transpose(1, 2) @ S_inv_square_R
+        merged = torch.cat([Sigma[:, :, 0], Sigma[:, 1:, 1], Sigma[:, 2:, 2], B.squeeze(), C], dim=1)
+
+        return merged
 
     @torch.no_grad()
     def compute_3D_filter(self, cameras):

scripts/show_nerfsynthetic.py

@@ -0,0 +1,33 @@
+import os
+import json
+import numpy as np
+import trimesh
+
+scenes = ["chair", "drums", "ficus", "hotdog", "lego", "materials", "mic", "ship"]
+
+output_dirs = ["exp_nerf_synthetic/release"]
+
+results = []
+for scene in scenes:
+    print(scene,)
+    for output in output_dirs:
+        json_file = f"{output}/{scene}/results.json"
+        data = json.load(open(json_file))
+        data = data['ours_30000'] if 'ours_30000' in data else data['ours_7000']
+
+        iteration = "30K iter: "
+        point_cloud_file = f"{output}/{scene}/point_cloud/iteration_30000/point_cloud.ply"
+        if not os.path.exists(point_cloud_file):
+            point_cloud_file = f"{output}/{scene}/point_cloud/iteration_7000/point_cloud.ply"
+            iteration = "7K iter: "
+        print(iteration, data.values(), trimesh.load(point_cloud_file).vertices.shape)
+        results.append(data['PSNR'])
+
+results = np.array(results).reshape(8, -1)
+
+print("===================")
+print("PSNR:")
+print(results)
+print("mean:")
+print(results.mean(axis=0))
+