vocab_tree_builder always builds a 128-d index, so trees can't be trained on LoMa / ALIKED descriptors #4918

Describe the bug
vocab_tree_builder always builds a 128-dimensional visual index, whatever descriptors are in the database. With LoMa-B features (256-d) it loads every descriptor and then aborts:

visual_index.cc:435 Check failed: descriptors.data.cols() == kDescDim (256 vs. 128)

So a vocabulary tree can't be trained on LoMa (or any non-128-d) features.

To Reproduce

1. Build COLMAP with ONNX.
2. colmap feature_extractor --database_path db.db --image_path images --FeatureExtraction.type LOMA_B
3. colmap vocab_tree_builder --database_path db.db --vocab_tree_path tree.bin --num_visual_words 4096
4. See the error above, raised from FaissVisualIndex<128, 64>::Build.

Expected behavior
The tree is built with the descriptors' own dimension (256 → embedding 64 for LoMa-B), as for SIFT, and can then be used by vocab_tree_matcher.

Screenshots
N/A

Environment:

• OS: Ubuntu 22.04 (WSL2), CUDA 12.9, RTX 5090
• COLMAP Version: 4.2.0 (be5e291) and 4.2.1 (bd1fcf6); the same code is in main at 9371f7b
• Capture Device: GoPro HERO13 Black (any images reproduce it)

Additional context
src/colmap/exe/vocab_tree.cc:65 calls retrieval::VisualIndex::Create() with no arguments. It therefore gets the defaults desc_dim = 128, embedding_dim = 64 (visual_index.h:83). The rest of the pipeline already handles other sizes:

• Create() implements 256/64 and 32/16.
• Trees store their descriptor size, embedding size and feature type, and Read() restores them.
• vocab_tree_matcher checks the feature type.
• A pretrained LoMa-B tree is shipped.

Only the builder is stuck at 128.

Suggested fix: size the index from the loaded descriptors.

const int desc_dim = descriptors.data.cols();
const int embedding_dim = desc_dim == 32 ? 16 : 64; // the pairs Create() implements
auto visual_index = retrieval::VisualIndex::Create(desc_dim, embedding_dim);

An optional --embedding_dim (0 = derive) would cover dimension/embedding pairs added later. SIFT is unchanged (128 → 64). With this patch on 4.2.1, a per-scan LoMa-B tree builds (Descriptor dimension 256 (LOMA_B), embedding dimension 64, 4096 words), and matching and global_mapper run normally. Happy to open a PR.