On model_splitter: After spliting the model, the output reconstruction is the same as the input reconstruction

[Serenitysmk]

Describe the bug

This may be a bug, or maybe I didn't use understand the usage of the command model_splitter .

I try to use split the reconstruction into 4 parts using the command like this:

colmap model_splitter \
--input_path ~my_input_path \
--output_path ~my_output_path \
--split_type parts \
--split_params "4"

And, the console message is:

==============================================================================
Splitting sparse model
==============================================================================

 => Using "parts" split type

Computing bound_coords
----------------------

Applying split and writing reconstructions
------------------------------------------
 => Splitting to 4 parts
Writing reconstruction 3 with 557 images, 145957 points, and 33.86% area coverage
Writing reconstruction 0 with 401 images, 443576 points, and 51.17% area coverage
Writing reconstruction 2 with 1105 images, 712636 points, and 40.77% area coverage
Writing reconstruction 1 with 1137 images, 1099260 points, and 50.82% area coverage
Elapsed time: 0.450 [minutes]

It looks like the model is divided into 4 parts. However, when writing the 4 sub-models into the output directory, I saw that the sub-models are exactly the same as the original input reconstruction.

I don't know if this is a bug or if I didn't use it properly. After some searching, I didn't find similar issues.

So I looked into the code at line 959 of ~src/exe/model.cc:

  auto SplitReconstruction = [&](const int idx) {
    Reconstruction tile_recon = reconstruction.Crop(bounds[idx]);
    // calculate area covered by model as proportion of box area
    auto bbox_extent = bounds[idx].second - bounds[idx].first;
    auto model_bbox = tile_recon.ComputeBoundingBox();
    auto model_extent = model_bbox.second - model_bbox.first;
    double area_ratio =
        (model_extent(0) * model_extent(1)) / (bbox_extent(0) * bbox_extent(1));
    int tile_num_points = tile_recon.NumPoints3D();

    std::string name = use_tile_keys ? tile_keys[idx] : std::to_string(idx);
    const bool include_tile =
        area_ratio >= min_area_ratio &&       //
        tile_num_points >= min_num_points &&  //
        tile_recon.NumRegImages() >= static_cast<size_t>(min_reg_images);

    if (include_tile) {
      std::cout << StringPrintf(
                       "Writing reconstruction %s with %d images, %d points, "
                       "and %.2f%% area coverage",
                       name.c_str(), tile_recon.NumRegImages(), tile_num_points,
                       100.0 * area_ratio)
                << std::endl;
      const std::string reconstruction_path = JoinPaths(output_path, name);
      CreateDirIfNotExists(reconstruction_path);
      reconstruction.Write(reconstruction_path);
      WriteBoundingBox(reconstruction_path, bounds[idx]);
      WriteBoundingBox(reconstruction_path, exact_bounds[idx], "_exact");

    } else {
      std::cout << StringPrintf(
                       "Skipping reconstruction %s with %d images, %d points, "
                       "and %.2f%% area coverage",
                       name.c_str(), tile_recon.NumRegImages(), tile_num_points,
                       100.0 * area_ratio)
                << std::endl;
    }
  };

I saw that the input reconstruction is splitted and cropped and saved into tile_recon, but calling reconstruction.Write(reconstruction_path) to the output path, which is the original reconstruction.

Am I correct? Or if I understood somewhere wrong?

[Serenitysmk]

Thanks for the fix. I am closing the issue.