omvg fix

README.md

@@ -5,7 +5,7 @@ A free and open source implementation of 3D [gaussian splatting](https://www.you
 <img src="https://github.com/pierotofy/OpenSplat/assets/1951843/c9327c7c-31ad-402d-a5a5-04f7602ca5f5" width="49%" />
 <img src="https://github.com/pierotofy/OpenSplat/assets/1951843/eba4ae75-2c88-4c9e-a66b-608b574d085f" width="49%" />
 
-OpenSplat takes camera poses + sparse points in [COLMAP](https://colmap.github.io/), [OpenSfM](https://github.com/mapillary/OpenSfM), [ODM](https://github.com/OpenDroneMap/ODM) or [nerfstudio](https://docs.nerf.studio/quickstart/custom_dataset.html) project format and computes a [scene file](https://drive.google.com/file/d/12lmvVWpFlFPL6nxl2e2d-4u4a31RCSKT/view?usp=sharing) (.ply or .splat) that can be later imported for [viewing](https://antimatter15.com/splat/?url=https://splat.uav4geo.com/banana.splat), editing and rendering in other [software](https://github.com/MrNeRF/awesome-3D-gaussian-splatting?tab=readme-ov-file#open-source-implementations).
+OpenSplat takes camera poses + sparse points in [COLMAP](https://colmap.github.io/), [OpenSfM](https://github.com/mapillary/OpenSfM), [ODM](https://github.com/OpenDroneMap/ODM) [OpenMVG](https://github.com/OpenMVG/OpenMVG) or [nerfstudio](https://docs.nerf.studio/quickstart/custom_dataset.html) project format and computes a [scene file](https://drive.google.com/file/d/12lmvVWpFlFPL6nxl2e2d-4u4a31RCSKT/view?usp=sharing) (.ply or .splat) that can be later imported for [viewing](https://antimatter15.com/splat/?url=https://splat.uav4geo.com/banana.splat), editing and rendering in other [software](https://github.com/MrNeRF/awesome-3D-gaussian-splatting?tab=readme-ov-file#open-source-implementations).
 
 Graphics card recommended, but not required! OpenSplat runs the fastest on NVIDIA, AMD and Apple (Metal) GPUs, but can also run entirely on the CPU (~100x slower).
 

model.hpp

@@ -4,6 +4,7 @@
 #include <iostream>
 #include <torch/torch.h>
 #include <torch/csrc/api/include/torch/version.h>
+#include <ATen/ATen.h>
 #include "nerfstudio.hpp"
 #include "kdtree_tensor.hpp"
 #include "spherical_harmonics.hpp"

openmvg.cpp

@@ -5,6 +5,7 @@
 #include "tensor_math.hpp"
 #include <stdexcept>
 #include <string>
+#include <bits/stdc++.h>
 
 namespace fs = std::filesystem;
 using json = nlohmann::json;
@@ -176,18 +177,43 @@ bool read_poses(const json& data, std::unordered_map<uint32_t, Pose> &poses){
             }
         }
     */
+
+    /*
+        OpenMVG rotation data is stored as columns, must convert to rows
+    */
+
     auto scene_poses = data["extrinsics"];
     for(auto item: scene_poses){
 
         std::uint32_t id = item["key"].get<std::uint32_t>();
         auto value = item["value"];
 
         Pose pose;
+        std::vector<float> omvg_rotation;
+
         for(auto row: value["rotation"]){
             for(auto r: row){
-                pose.rotation.push_back( r.get<float>() );
+                omvg_rotation.push_back( r.get<float>() );
             }
         }
+
+        pose.rotation = std::vector<float>(9);
+
+        // convert cols to rows
+
+        pose.rotation[0] = omvg_rotation[0];
+        pose.rotation[1] = omvg_rotation[3];
+        pose.rotation[2] = omvg_rotation[6];
+
+        pose.rotation[3] = omvg_rotation[1];
+        pose.rotation[4] = omvg_rotation[4];
+        pose.rotation[5] = omvg_rotation[7];
+
+        pose.rotation[6] = omvg_rotation[2];
+        pose.rotation[7] = omvg_rotation[5];
+        pose.rotation[8] = omvg_rotation[8];
+
+
         for(auto c: value["center"]){
             pose.center.push_back( c.get<float>() );
         }
@@ -256,24 +282,26 @@ InputData inputDataFromOpenMVG(const std::string &projectRoot){
     for (const auto &p : poses){
         std::uint32_t pose_id = p.first;
         Pose pose = p.second;
+
+        torch::Tensor R = torch::from_blob(pose.rotation.data(), {static_cast<long>(pose.rotation.size())}, torch::kFloat32);
+        R = R.reshape({3, 3});
 
-        torch::Tensor rotation = rodriguesToRotation(torch::from_blob(pose.rotation.data(), {static_cast<long>(pose.rotation.size())}, torch::kFloat32));
-        torch::Tensor translation = torch::from_blob(pose.center.data(), {static_cast<long>(pose.center.size())}, torch::kFloat32);
-        torch::Tensor w2c = torch::eye(4, torch::kFloat32);
-        w2c.index_put_({Slice(None, 3), Slice(None, 3)}, rotation);
-        w2c.index_put_({Slice(None, 3), Slice(3,4)}, translation.reshape({3, 1}));
-
-        unorientedPoses[i] = torch::linalg::inv(w2c);
-
-        std::cout << pose_id << " " << i;
+        torch::Tensor T = torch::from_blob(pose.center.data(), {static_cast<long>(pose.center.size())}, torch::kFloat32);
+        T = T.reshape({3, 1});
+
+        torch::Tensor Rinv = R.transpose(0, 1);
+        torch::Tensor Tinv = torch::matmul(-Rinv, T);
 
         // because the maps are unordered, need this to keep track of which pose in the tensor is the pose we need
         pose_indexes[pose_id] = i;
 
-        // Convert OpenSfM's camera CRS (OpenCV) to OpenGL
+        unorientedPoses[i].index_put_({Slice(None, 3), Slice(None, 3)}, R);
+        unorientedPoses[i].index_put_({Slice(None, 3), Slice(3, 4)}, Tinv);
+        unorientedPoses[i][3][3] = 1.0f;
+
+        // Convert OpenMVG's camera CRS (OpenCV) to OpenGL
         unorientedPoses[i].index_put_({Slice(0, 3), Slice(1,3)}, unorientedPoses[i].index({Slice(0, 3), Slice(1,3)}) * -1.0f);
         i++;
-
     }
 
     std::cout << "  " << std::endl;
@@ -296,11 +324,8 @@ InputData inputDataFromOpenMVG(const std::string &projectRoot){
         fs::path thisRoot(image_root_path);
         fs::path image_path =  thisRoot/ v.s_Img_path;
 
-
         std::uint32_t current_pose = pose_indexes.at(v.id_pose);
 
-        std::cout << view_id << " " << v.s_Img_path << " pose " << current_pose << std::endl;
-
         float normalizer = static_cast<float>((std::max)(intrinsic.width, intrinsic.height));
         ret.cameras.emplace_back(Camera(intrinsic.width, intrinsic.height, 
                             static_cast<float>(intrinsic.fx * normalizer), static_cast<float>(intrinsic.fy * normalizer),