Calculate localization error using sfm_gt as reference model

[davidng82]

I ran COLMAP successfully with my RGB data. My dataset includes synchronized RGB, depth, and IMU data for each timestamp. After running COLMAP, I used the script create_gt_sfm.py to generate a ground-truth model (sfm_gt) using depth information.

Now, I want to calculate localization error. Here's what I’ve done so far:

1. Split RGB data
   I split the RGB images into an 80% training set and 20% test set.

2. Feature extraction (on training set)
   Local features:
   python -m hloc.extract_features
   --image_dir images_train
   --export_dir outputs/train_hloc
   --conf superpoint_max

Global features:
python -m hloc.extract_features
--image_dir images_train
--export_dir outputs/train_hloc
--conf netvlad

3. Image retrieval (get pairs)
   python -m hloc.pairs_from_retrieval
   --descriptors outputs/train_hloc/global-feats-netvlad.h5
   --output outputs/train_hloc/pairs-train-netvlad.txt
   --num_matched 20

4. Match features
   python -m hloc.match_features
   --pairs outputs/train_hloc/pairs-train-netvlad.txt
   --features feats-superpoint-n4096-r1024
   --export_dir outputs/train_hloc
   --conf superglue

5. Problem when running triangulation
   python -m hloc.triangulation
   --sfm_dir outputs/train_hloc/sfm_localized
   --reference_sfm_model outputs/sfm_gt
   --image_dir images_train
   --pairs outputs/train_hloc/pairs-train-netvlad.txt
   --features outputs/train_hloc/features.h5
   --matches outputs/train_hloc/matches.h5

But I get an error like this: KeyError: "Unable to synchronously open object (object 'XYZ.png' doesn't exist)".
I assume this is because the sfm_gt was generated from the full RGB dataset, while I am now trying to triangulate using only the 80% training subset (images_train).

Questions:
For the --reference_sfm_model argument, is it okay to use sfm_gt, even though it includes all RGB images?
If not, should I generate a new sfm_gt that uses only the images_train subset?
Is there a way to skip the missing-image check or restrict triangulation to the intersecting images?

Any suggestions would be greatly appreciated!