Generate localization map

XRLocalization map includes four files: images.bin, points3D.bin, features.bin and database.bin. In this part, we introduce how to generate this map.

The format of images.bin and points3D.bin is the same as COLMAP format. features.bin records 3D descriptor for all 3D point in points3D.bin. Every 3D descriptor is represented as the mean of all it corresponding 2D local features. database.bin records image global features for all image recorded in images.bin.

The prerequisite is re-triangulation. Two type of re-triangulation results are supported.

• Re-triangulation by xrloc, requiring prerequisites including: images.bin, points3D.bin and features.bin
• Re-triangulation by hloc, requiring prerequisites including: images.bin, points3D.bin and feats-xxxx.h5

Step 0 Generate images.bin point3Ds.bin features.bin

xrloc:

python3 tools/loc_convert_reconstruction.py \
    --feature_path /path/to/features.bin \
    --model_path /path/to/include/images.bin/and/points3Ds.bin/directory \
    --output_path /path/to/map/directory

This would produce three new file images.bin, points3D.bin, features.bin in the path /path/to/map/directory.

hloc:

python3 tools/loc_convert_reconstruction.py \
    --feature_path /path/to/feats-xxxx.h5 \
    --model_path /path/to/include/images.bin/and/points3Ds.bin/directory \
    --output_path /path/to/map/directory

The only difference is at --feature_path /path/to/feats-xxxx.h5. The output is the same as xrloc.

Step 1 Generate database.bin

python3 tools/ir_create_database.py \
   --image_dir /path/to/image_dir \
   --image_bin_path /path/to/step0/map/directory/images.bin \
   --databse_path /data/to/database.bin \
   --extractor netvlad

This would extract netvlad feature for all images in image_dir and save as database.bin.

Finally, merge the output from two steps as localization map.