feature-retrieval
Feature extraction, retrieval descriptors, matching, and HDF5 feature/match formats for hloc.
- License 许可
- Apache-2.0
- 最近更新
- 2026-09-03
- Stars
- ★ 351
- Forks
- 27
01 前言 Front Matter
- name
- feature-retrieval
- license
- Apache 2.0
- metadata
- { "disco-role": "operating" }
- description
- Feature extraction, retrieval descriptors, matching, and HDF5 feature/match formats for hloc.
- disable-model-invocation
- true
02 资源文件 4
- 参考文档references/api-and-cli.md16.6 KB
- 参考文档references/hdf5-feature-match-formats.md8.8 KB
- 参考文档references/troubleshooting.md10.5 KB
- 脚本scripts/inspect_hloc_configs.py7.0 KB
03 SKILL.md 正文
Feature Retrieval for Hierarchical-Localization
Use this sub-skill when a task needs hloc feature files, global retrieval descriptor files, retrieval pairs, sparse match files, dense LoFTR-derived match files, or guidance for choosing built-in extractor/matcher configurations.
Route here for
- Local feature extraction with
hloc.extract_features.main,ImageDataset, andextract_features.confs. - Global image retrieval descriptors from the extractor configs
dir,netvlad,openibl, ormegaloc, followed byhloc.pairs_from_retrieval.main. - Sparse matching with
hloc.match_features.main, including SuperGlue, LightGlue, nearest-neighbor, and AdaLAM config families. - Dense/semi-dense matching with
hloc.match_dense.mainand LoFTR config families. - HDF5 feature, global descriptor, and match file naming/schema diagnostics.
Route elsewhere
- SfM reconstruction, triangulation, pair generation beyond retrieval descriptors, and localization execution: use
../mapping-localization/. - Benchmark/dataset pipeline scripts and large dataset preparation: use
../dataset-pipelines/. - Implementing custom extractors/matchers or exporting external feature/match files: use
../custom-interop/.
Start with these references
- Read
references/api-and-cli.mdfor verified signatures, CLI options, exact built-in config names, output naming, and config-family selection. - Read
references/hdf5-feature-match-formats.mdbefore reading or writing.h5feature, global descriptor, or match files. - Read
references/troubleshooting.mdwhen extraction, retrieval, matching, model download, CUDA/CPU, or HDF5 schema errors occur. - Run
scripts/inspect_hloc_configs.pyin the user's active Python environment to list available installed configs without running neural models.
Minimal safe patterns
Extract local features
from pathlib import Path
from hloc import extract_features
images = Path("images")
outputs = Path("outputs")
feature_conf = extract_features.confs["superpoint_aachen"]
feature_path = extract_features.main(feature_conf, images, outputs)
# Usually: outputs / "feats-superpoint-n4096-r1024.h5"
Extract global descriptors and produce retrieval pairs
from pathlib import Path
from hloc import extract_features, pairs_from_retrieval
images = Path("images")
outputs = Path("outputs")
retrieval_conf = extract_features.confs["netvlad"]
descriptor_path = extract_features.main(retrieval_conf, images, outputs)
pairs = outputs / "pairs-netvlad.txt"
pairs_from_retrieval.main(descriptor_path, pairs, num_matched=20)
Match sparse features for a pairs file
from pathlib import Path
from hloc import extract_features, match_features
outputs = Path("outputs")
pairs = outputs / "pairs-netvlad.txt"
feature_conf = extract_features.confs["superpoint_aachen"]
matcher_conf = match_features.confs["superglue"]
feature_path = outputs / f"{feature_conf['output']}.h5"
match_path = match_features.main(matcher_conf, pairs, feature_conf["output"], outputs)
# Usually: outputs / "feats-superpoint-n4096-r1024_matches-superglue_pairs-netvlad.h5"
Match with dense LoFTR
from pathlib import Path
from hloc import match_dense
images = Path("images")
outputs = Path("outputs")
pairs = outputs / "pairs-netvlad.txt"
dense_conf = match_dense.confs["loftr_aachen"]
features_path, matches_path = match_dense.main(dense_conf, pairs, images, export_dir=outputs)
Do not assume a feature or match path from memory: compute it from the selected config and pair-file stem, or pass explicit paths via the Python API.
04 同仓库更多 Skills
passnet-triton-opt
Make a single PassNet Triton kernel fast AND numerically faithful: performance model (when a replacement can win at all), block/grid/warp tuning, autotune policy, launch overhead, and per-op numeric recipes to pass the dtype baseline tolerances. Use when a pass matches and is correct but speedup below expected, or when correctness fails by small numeric margins.
passnet-solve
END-TO-END playbook for solving one PassNet sample: analyze the computation graphs, decide the optimization strategy (what to fuse, what to replace, what to leave alone), drive the iteration loop, and maximize the sample score. This is the ENTRY skill — invoke it first for any PassNet optimization task; it tells you when to use passnet-pattern-fusion, passnet-triton-opt and passnet-feedback.
passnet-skill
PassNet GPU kernel optimization via compiler passes. Design and implement Triton-based optimization passes, create pass files under ./pass_dir/, self-evaluate with pass_evaluator, and iterate to maximize GPU speedup.
passnet-pattern-fusion
Author PassNet pass files: write patterns that actually MATCH the FX graph, pick fusion regions (which ops to absorb into one Triton kernel), and structure multi-pass submissions with the mandatory shared-dispatch architecture. Use when creating or fixing pattern()/replacement_args()/replacement_func() files, when a pass "failed to match", or when deciding how to fuse multiple kernels/ops.
passnet-orchestrate
Round/batch planning and decision support for PassNet graph optimization. Use for multi-sample triage, worker allocation, eval-budget policy, and post-eval keep/revert decisions. For solving one concrete sample end to end, use passnet-solve as the entry skill; consult this skill only when you need broader planning or a specific decision gate.
passnet-feedback
Fast, deterministic feedback for PassNet work: pre-flight pattern/match verification WITHOUT burning a GPU evaluation, per-node bottleneck analysis of a sample's graphs, and eval-log parsing into per-variant status + estimated score + failure classification. Use BEFORE every GPU evaluation (check_pattern), at the START of a sample (analyze_graph), and AFTER every evaluation (parse_eval_log).