Generation of `model_train/sparse/points3D.ply`

[HsiehTPE]

Question about how model_train sparse model is constructed

Dear dev,

I am currently building my own dataset following the Nerfbusters format and am trying to understand how the model_train sparse model is generated.

After inspecting the binary files under the colmap/ directory, I confirmed that the full sparse model (cameras / images / points3D) appears to be reconstructed using both train and test images.

However, I noticed something interesting:

• The images.txt inside model_train appears to be identical to the one in colmap/.
• But the points3D file in model_train differs from the one in colmap/.

This raises a few questions:

1. Is model_train constructed by re-triangulating only the training images?
2. Or is it derived from the full sparse model by filtering out 3D points that are not observed by training images?

I am trying to faithfully reproduce the same structure for my own dataset, and I want to understand how the difference in points3D between colmap/ and model_train/ should be interpreted.

Thank you very much for your clarification.

[minsu1206]

Hi. Thanks for your interest in our work.

We consider following:

Initialization of 3DGS with point cloud from both training viewpoints and test viewpoints doesn't make sense in the case of exploration task. Training viewpoints and test viewpoints are very distant. Using pointclouds from test viewpoints gives lots of geometric hints to 3DGS.

Thus, the process is

(1) Run COLMAP with all training viewpoints and test viewpoints : this is necessary for estimation of precise camera poses
(2) Re-run COLMAP with only training viewpoints (or only test viewpoints) : this is to obtain precise point clouds, avoiding mentioned issue.
(2)-2 : To be specific, (I'll take an example case - using only training viewpoints)
a. re-write images.txt without test viewpoints ; only using training viewpoints
b. run feature extraction
c. run feature matching
d. run point triangulation

Feel free to ask
Thanks.

[HsiehTPE]

Dear dev,

Thank you very much for your detailed explanation and guidance. I really appreciate your help.

I will carefully follow the procedure you described and try to reproduce the results as soon as possible. Once I obtain the results, I will share my findings and observations here in this issue for reference.

In addition to this question, I also have another inquiry related to prompt generation for the diffusion model in the project. I will open a separate issue later to keep the discussion organized. I would greatly appreciate it if you could take a look when you have time.

Thanks again for your support and for maintaining this excellent work.

[minsu1206]

I’m glad that helped.

After finishing my on-going another project, I have a plan to release some codes related to dataset pre-processing - Sorry for being late. I need to re-arrange my codes...

For prompt generation, You can refer to data_preprocess/gpt_batch_scannetpp_example.py. This may be easy to understand.

Feel free to ask or open another git issue.
Thanks.

[HsiehTPE]

Hi,

I am trying to reproduce the COLMAP workflow used in your project and encountered an issue related to image_id consistency.

---

Problem

When performing subset triangulation (e.g., train-only) using fixed poses, I observed that:

• The database.db contains image entries with certain image_id -> image_name mappings.
• The filtered images.txt (derived from the full reconstruction) also contains IMAGE_ID -> NAME.

However, during point_triangulator, COLMAP aborts with:

Check failed: existing_image.Name() == image.second.Name()

This suggests that:

• The image_id in database.db
• And the IMAGE_ID in images.txt

do not correspond to the same image names.

---

My Current Workflow

1. Run COLMAP on all images (train + test) → obtain stable poses.
2. Convert model to TXT.
3. Filter images.txt to keep only train images.
4. Run point_triangulator using:
   • --database_path (from full reconstruction)
   • --input_path (filtered model)

This leads to the image_id mismatch error.

---

Questions

1. Is it required that database.db and images.txt contain identical image_id mappings?
2. When performing subset triangulation, how did you ensure consistency between:
   • database image entries
   • filtered model images.txt
3. Which version of COLMAP did you use in your experiments?

I am currently using COLMAP 3.8.

Any clarification on the intended workflow would be greatly appreciated.

Thank you!