HTL-ReID

HTL-ReID is a research codebase for robust multi-modal object re-identification with RGB / NIR / TIR inputs.

Mainline

The current evidence-backed path is A2: shared ViT-B/16, hierarchical token selection (HS), FACSS token filtering, and quality-aware selection/fusion weights (QAWF).

A3-A5 overlays, AGF/TPM fusion, modality adapters, part branches, and auxiliary loss/full branches are kept only for archived ablations or negative evidence. They are not the current paper-method path. HSL is not part of this model path.

Requirements

pip install -r requirements.txt

pytorch_wavelets is vendored under ./pytorch_wavelets; do not install it separately.

Set DATASETS.ROOT_DIR and MODEL.PRETRAIN_PATH_T in the dataset config, or override them from the command line.

Training

python train_net.py --config_file configs/RGBNT201/default.yml \
    --config_file configs/RGBNT201/ablations/chain20/a2_quality.yml \
    DATASETS.ROOT_DIR /path/to/datasets \
    MODEL.PRETRAIN_PATH_T /path/to/pretrained_vit.pth

The A2 overlay sets SOLVER.TRAIN_EPOCHS: 20 while keeping SOLVER.MAX_EPOCHS: 120, so the cosine schedule still follows the 120-epoch contract.

Evaluation

python test_net.py --config_file configs/RGBNT201/default.yml \
    --config_file configs/RGBNT201/ablations/chain20/a2_quality.yml \
    TEST.WEIGHT /path/to/checkpoint.pth

TEST.RE_RANKING is enabled by default in the dataset configs.

Smoke Test

python test_pipeline.py

The smoke test checks config merging, scheduler semantics, 3-modal and 2-modal forward/backward passes, save/load, and ablation switches without real datasets.

License

This project is released under the terms of the LICENSE file in this repository.