Reproduce

The whole pipeline is one installable package, wildmatch, with one command per step. Each step reads the previous step's outputs from the storage layout set by the path profile.

Step	Command	What it does
1. Pair mining	wildmatch mine	Builds feature caches, matches every training image against the others with the pretrained matcher, and writes the positive and negative pools (strong-match indices).
2. Matcher fine-tuning	wildmatch finetune-matcher	Samples triplets from the pools and fine-tunes the matching module (LoMa or RDD-LightGlue) with the triplet margin loss on the relaxed score; saves per-epoch checkpoints.
3. Evaluation	wildmatch evaluate	Runs the probes: cosine retrieval, WildFusion, default and fine-tuned matchers over the MegaDescriptor-L candidate list, and the classifier baselines; exports tables and figures.

Note

The commands shown are the ones used for the manuscript, with machine-specific paths replaced by <placeholders>. wildmatch <command> --help lists every option. The Slurm scripts in slurm/ target one specific cluster; adjust partitions and the environment location for yours.

Few-shot views run the same three steps on a reduced training split; they are an extension and not part of the paper.

The repository's guides go further than these pages: installation, dataset preparation for every registry dataset, configuration of every method, and experiments (sweeps, outputs, tables and figures).

Environment

One environment with pinned dependencies (Python 3.12) runs all three steps:

uv sync --extra cu126 --extra matchers --extra train --group dev   # or --extra cpu
source .venv/bin/activate

matchers adds the local matchers through Vismatch, pinned to a fixed commit, which downloads pretrained matcher weights on first use; train adds what mining and fine-tuning need (accelerate and LoMa, pinned). The RDD code used for mining and fine-tuning ships inside the package. Evaluation alone does not need train.

Data layout

Every location comes from a path profile: data_root for the datasets, checkpoint_root for feature caches and fine-tuned checkpoints, and external.mining_outputs for the mined indices. The default profile reads them from environment variables (WILDMATCH_DATA_ROOT, WILDMATCH_CHECKPOINT_ROOT, WILDMATCH_MINING_OUTPUTS), and the pretrained weights the same way (WILDMATCH_LOMA_WEIGHTS for loma_B.pt, WILDMATCH_RDD_WEIGHTS_DIR for the folder holding RDD-v2.pth and RDD_lg-v2.pth). Datasets are entries of the dataset registry, selected by key (nyala, salamander, czechlynx_closed, ...).

Each dataset is a root directory with an images/ or masked_images/ folder and a metadata CSV holding the image path, identity, split and, where applicable, a COCO-RLE mask. Mining and fine-tuning read a symlink view with one folder per individual and collection built from that metadata; evaluation reads the metadata directly. Feature caches and run artifacts are keyed by content hashes of images, metadata and checkpoints, so a changed input can never reuse stale features.

Shared settings

Setting	Value
Image long side	512 px
Keypoints per image	up to 512
Pairs per anchor	5 positives, 5 hard negatives
Training	300 epochs, AdamW \(10^{-5}\), effective batch 32, margin 0.5
Reported checkpoint	epoch 299, no selection
Candidate budgets	10, 50, 100, 250 (default), 500, 1000
Metrics	Top-5 accuracy, balanced Top-1 accuracy