Instance / Semantic Segmentation

[abrahamezzeddine]

I’m currently working with COLMAP, and I love it!

My plan is to perform object segmentation on the same set of images.

My goal is to integrate these two processes. First, I will use COLMAP to create a 3D sparse/dense model from the original, unsegmented images. The segmentation model will on the other hand process same set of images to extract specific object pixel coordinates.

The next step involves a comparative analysis where I’ll match the pixel coordinates from the segmented images with the pixels detected by COLMAP and cross reference them with the point cloud. This will allow me to annotate the COLMAP-generated 3D point cloud model with semantic labels based on the segmentation data.

I’m particularly interested in understanding how to efficiently extract the detected points and their corresponding pixel coordinates from COLMAP for this purpose.

I hope my task is explained in an easy way. Thank you. Have a good day.

[jytime]

you can check this document:

https://colmap.github.io/format.html

What you want is saved in points3D.txt and images.txt

[sarlinpe]

Or just pycolmap to inspect the resulting COLMAP model and fuse the semantic information of the different observations ofr each 3D point: https://github.com/colmap/colmap/tree/main/pycolmap#reconstruction-object

[lucascassiano]

@sarlinpe any examples on how to do it?

I would love to run a GroundedSAM over the input images and re-project on the 3D point cloud

[sarlinpe]

For regular semantic segmentation:

import pycolmap
rec = pycolmap.Reconstruction("path/to/sfm/folder")
p3d_to_observations = defaultdict(list)
for image in rec.images.values():
  segmentation = get_segmentation(image.name)  # class probabilities, (H, W, C)
  for p2d_id in image.get_valid_point2D_ids():
    p2d = image.points2D[p2d_id]
    xy = p2d.xy
    ij = tuple(xy.astype(int))
    semantics = segmentation[ij]
    p3d_to_observations[p2d.point3D_id].append(semantics)
p3d_to_fused = {k: np.mean(v, 0) for k, v in p3d_to_observations.items()}  # fuse the semantic probs
p3d_to_class = {k: np.argmax(v) for k, v in p3d_to_fused.items()}  # pick top prob

[abrahamezzeddine]

@sarlinpe any examples on how to do it?

I would love to run a GroundedSAM over the input images and re-project on the 3D point cloud

Any success with this?