Feat: built-in rad support

CMakeLists.txt

@@ -9,8 +9,6 @@ set(OPENSPLAT_BUILD_VISUALIZER OFF CACHE BOOL "Build visualizer application")
 set(OPENSPLAT_USE_FAST_MATH OFF CACHE BOOL "Enable fast math optimizations for GPU kernels (-use_fast_math / -ffast-math)")
 set(OPENSPLAT_USE_PCH ON CACHE BOOL "Use precompiled headers to speed up compilation")
 
-set(FETCH_DEPENDENCIES ON CACHE BOOL "Fetch additional dependencies from the Internet during configuration")
-
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR})
 
 # Read version
@@ -53,41 +51,59 @@ include(FetchContent)
 
 set(NANOFLANN_BUILD_EXAMPLES OFF)
 set(NANOFLANN_BUILD_TESTS OFF)
-set(SPZ_BUILD_PYTHON_BINDINGS OFF)
-set(SPZ_BUILD_TOOLS OFF)
-set(SPZ_BUILD_EXTENSIONS OFF)
-set(SPZ_BUILD_WASM OFF)
-set(SPZ_SHOULD_INSTALL OFF)
+set(ZLIB_BUILD_TESTING OFF)
+set(ZLIB_BUILD_SHARED OFF)
+set(ZLIB_INSTALL OFF)
 
-if(FETCH_DEPENDENCIES)
+find_package(nlohmann_json QUIET)
+if(NOT nlohmann_json_FOUND)
+    message(STATUS "nlohmann_json not found, fetching it")
     FetchContent_Declare(nlohmann_json
         URL https://github.com/nlohmann/json/archive/refs/tags/v3.11.3.zip
     )
+    FetchContent_MakeAvailable(nlohmann_json)
+endif()
+
+find_package(nanoflann QUIET)
+if(NOT nanoflann_FOUND)
+    message(STATUS "nanoflann not found, fetching it")
     FetchContent_Declare(nanoflann
         URL https://github.com/jlblancoc/nanoflann/archive/refs/tags/v1.5.5.zip
     )
+    FetchContent_MakeAvailable(nanoflann)
+endif()
+
+find_package(cxxopts QUIET)
+if(NOT cxxopts_FOUND)
+    message(STATUS "cxxopts not found, fetching it")
     FetchContent_Declare(cxxopts
         URL https://github.com/jarro2783/cxxopts/archive/refs/tags/v3.2.0.zip
     )
-    FetchContent_Declare(spz
-        URL https://github.com/nianticlabs/spz/archive/affd0ecea7fbb4c265ee119475af7ee5b2997482.zip
-    )
-    FetchContent_MakeAvailable(nlohmann_json nanoflann cxxopts spz)
+    FetchContent_MakeAvailable(cxxopts)
+endif()
+
+find_package(ZLIB QUIET)
+if(ZLIB_FOUND)
+    set(ZLIB_LIB ZLIB::ZLIB)
 else()
-    find_package(nlohmann_json)
-    find_package(nanoflann)
-    find_package(cxxopts)
-    find_package(spz)
+    message(STATUS "ZLIB not found, fetching it")
+    FetchContent_Declare(zlib
+        URL https://github.com/pierotofy/OpenSplat/releases/download/v1.1.4/zlib-1.3.2.tar.gz
+    )
+    FetchContent_MakeAvailable(zlib)
+    set(ZLIB_LIB zlibstatic)
 endif()
 
+add_subdirectory(vendor/spz)
+
 if((GPU_RUNTIME STREQUAL "CUDA") OR (GPU_RUNTIME STREQUAL "HIP"))
-    if(FETCH_DEPENDENCIES)
+    find_package(glm QUIET)
+    if(NOT glm_FOUND)
+        message(STATUS "glm not found, fetching it")
         FetchContent_Declare(glm
             URL https://github.com/g-truc/glm/archive/refs/tags/1.0.1.zip
         )
         FetchContent_MakeAvailable(glm)
-    else()
-        find_package(glm)
     endif()
 endif()
 
@@ -265,7 +281,7 @@ target_include_directories(gsplat_cpu PRIVATE ${TORCH_INCLUDE_DIRS})
 set(OPENSPLAT_SRC_FILES opensplat.cpp point_io.cpp nerfstudio.cpp model.cpp
 kdtree_tensor.cpp spherical_harmonics.cpp cv_utils.cpp utils.cpp project_gaussians.cpp
 rasterize_gaussians.cpp ssim.cpp optim_scheduler.cpp colmap.cpp opensfm.cpp openmvg.cpp input_data.cpp
-tensor_math.cpp)
+tensor_math.cpp rad.cpp)
 
 if (OPENSPLAT_BUILD_VISUALIZER)
     if (Pangolin_FOUND)
@@ -302,6 +318,7 @@ target_link_libraries(opensplat PRIVATE
     nlohmann_json::nlohmann_json
     cxxopts::cxxopts
     nanoflann::nanoflann
+    ${ZLIB_LIB}
     spz::spz
 )
 if (NOT WIN32)

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), [ODX](https://github.com/WebODM/ODX),  [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, .splat, or .spz) 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), [ODX](https://github.com/WebODM/ODX),  [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, .splat, .spz, or .rad) 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.cpp

@@ -1,4 +1,5 @@
 #include <filesystem>
+#include <load-spz.h>
 #include "model.hpp"
 #include "constants.hpp"
 #include "splat-types.h"
@@ -8,7 +9,7 @@
 #include "tensor_math.hpp"
 #include "gsplat.hpp"
 #include "utils.hpp"
-#include <load-spz.h>
+#include "rad.hpp"
 
 #ifdef USE_MPS
 #include <torch/mps.h>
@@ -510,6 +511,14 @@ void Model::save(const std::string &filename, int step){
         savePly(filename, step);
         std::cout << "Wrote " << filename << std::endl;
     }
+    else if (extension == ".rad") {
+        if (saveRad(filename)) {
+            std::cout << "Wrote " << filename << std::endl;
+        }
+        else {
+            std::cerr << "Failed to write " << filename << ", aborting save." << std::endl;
+        }
+    }
     else {
         bool success = saveSpz(filename);
         if (success) {
@@ -621,14 +630,14 @@ bool Model::saveSpz(const std::string &filename){
 
 
     torch::Tensor meansCpu = keepCrs ? (means.cpu() / scale) + translation : means.cpu();
-    torch::Tensor scalesCpu = keepCrs ? (scales.cpu() / scale) : scales.cpu();
-    
+    torch::Tensor scalesCpu = keepCrs ? torch::log(torch::exp(scales.cpu()) / scale) : scales.cpu();
+
     torch::Tensor meansFlat = meansCpu.flatten();
     torch::Tensor scalesFlat = scalesCpu.flatten();
     torch::Tensor colorsFlat = featuresDc.cpu().flatten(); // raw DC coefficients
     torch::Tensor opacFlat = opacities.flatten().cpu();
-    torch::Tensor quatsFlat = quats.flatten().cpu();
-    torch::Tensor shRestFlat = featuresRest.cpu().transpose(1, 2).flatten();
+    torch::Tensor quatsFlat = torch::roll(quats.cpu(), -1, 1).flatten();
+    torch::Tensor shRestFlat = featuresRest.cpu().flatten();
 
     spz::GaussianCloud gaussians;
     gaussians.numPoints = meansCpu.size(0);
@@ -641,16 +650,41 @@ bool Model::saveSpz(const std::string &filename){
     gaussians.colors = tensor_to_vector<float>(colorsFlat);
     gaussians.sh = tensor_to_vector<float>(shRestFlat);
 
-    auto options = spz::PackOptions{
-        .version = 3,       // V4 available but not handled by many viewers
-        .from = spz::CoordinateSystem::RUB,
-        .sh1Bits = 6,
-        .shRestBits = 5
-    };
+    spz::PackOptions options;
+    options.version = 3;       // V4 available but not handled by many viewers
+    options.from = spz::CoordinateSystem::RUB;
+    options.sh1Bits = 6;
+    options.shRestBits = 5;
     bool success = spz::saveSpz(gaussians, options, filename);
     return success;
 }
 
+bool Model::saveRad(const std::string &filename){
+    size_t numPoints = means.size(0);
+
+    // Same value preparation as savePly; rad.cpp expects exactly the values
+    // a saved PLY would contain
+    torch::Tensor meansCpu = (keepCrs ? (means.cpu() / scale) + translation : means.cpu()).contiguous();
+    torch::Tensor featuresDcCpu = featuresDc.cpu().contiguous();
+    // featuresRest [N, K, 3] coefficient-major matches rad's SH layout; no transpose
+    torch::Tensor featuresRestCpu = featuresRest.cpu().contiguous();
+    torch::Tensor opacitiesCpu = opacities.cpu().contiguous();
+    torch::Tensor scalesCpu = (keepCrs ? torch::log(torch::exp(scales.cpu()) / scale) : scales.cpu()).contiguous();
+    torch::Tensor quatsCpu = quats.cpu().contiguous();
+
+    rad::SplatData data;
+    data.numPoints = numPoints;
+    data.numRestCoeffs = featuresRest.size(1);
+    data.means = tensor_to_vector<float>(meansCpu);
+    data.featuresDc = tensor_to_vector<float>(featuresDcCpu);
+    data.featuresRest = tensor_to_vector<float>(featuresRestCpu);
+    data.opacities = tensor_to_vector<float>(opacitiesCpu);
+    data.scales = tensor_to_vector<float>(scalesCpu);
+    data.quats = tensor_to_vector<float>(quatsCpu);
+
+    return rad::saveRad(filename, data);
+}
+
 void Model::saveDebugPly(const std::string &filename, int step){
     // A standard PLY
     std::ofstream o(filename, std::ios::binary);

model.hpp

@@ -73,6 +73,7 @@ struct Model{
   void savePly(const std::string &filename, int step);
   void saveSplat(const std::string &filename);
   bool saveSpz(const std::string &filename);
+  bool saveRad(const std::string &filename);
   void saveDebugPly(const std::string &filename, int step);
   int loadPly(const std::string &filename);
   torch::Tensor mainLoss(torch::Tensor &rgb, torch::Tensor &gt, float ssimWeight);

opensplat.cpp

@@ -160,7 +160,7 @@ int main(int argc, char *argv[]){
 
             if (step % displayStep == 0) {
                 const float percentage = static_cast<float>(step) / numIters;
-                std::cout << "Step " << step << ": " << mainLoss.item<float>() << " (" << floor(percentage * 100) << "%)" <<  std::endl;
+                std::cout << "Step " << step << ": " << mainLoss.item<float>() << " [" << floor(percentage * 100) << "%]" <<  std::endl;
             }
 
             model.optimizersStep();