How msm8976/whu-mars works
WHU-MARS provides a multispectral aerial-ground person ReID benchmark; CVPR26_UAD implements UAD training and retrieval evaluation. The main workflow configures a run, loads and samples modality-specific images, trains a ViT-based model with identity/metric losses and UAD alignment and prototype objectives, then evaluates retrieval rankings. The README documents ProCA and GPD; processor excerpts show their use. Some evaluation wiring is only partly sampled, so the graph omits unverified test-entry relationships.
Main entry points: train.py Training entry · make_dataloader.py Data loaders · processor.py Training processor · make_model.py ReID model
Components 18
Run setup
- Run configuration
CVPR26_UAD/config/defaults.py - Training entry
CVPR26_UAD/train.py
Data preparation
- WHU-MARS dataset
CVPR26_UAD/datasets/whu_mars.py - Image datasets
CVPR26_UAD/datasets/bases.py - Data loaders
CVPR26_UAD/datasets/make_dataloader.py - Identity samplers
CVPR26_UAD/datasets/sampler.py - Distributed sampler
CVPR26_UAD/datasets/sampler_ddp.py
Representation learning
- ReID model
CVPR26_UAD/model/make_model.py - Vision Transformer
CVPR26_UAD/model/backbones/vit_pytorch.py - Identity and metric losses
CVPR26_UAD/loss/make_loss.py - ProCA alignment
CVPR26_UAD/processor/processor.py - GPD prototypes
CVPR26_UAD/processor/processor.py - Training processor
CVPR26_UAD/processor/processor.py - Optimizer
CVPR26_UAD/solver/make_optimizer.py - Learning-rate schedule
CVPR26_UAD/solver/scheduler_factory.py
Retrieval evaluation
- Retrieval metrics
CVPR26_UAD/utils/metrics.py - Distance reranking
CVPR26_UAD/utils/reranking.py
Other components
- Researcher
Connections 17
- Researcher → Training entry: starts training
CVPR26_UAD/train.py - Training entry → Run configuration: loads settings
CVPR26_UAD/train.py - Training entry → Data loaders: builds loaders
CVPR26_UAD/train.py - Data loaders → WHU-MARS dataset: loads splits
CVPR26_UAD/datasets/make_dataloader.py - Data loaders → Image datasets: wraps samples
CVPR26_UAD/datasets/make_dataloader.py - Data loaders → Identity samplers: uses sampler
CVPR26_UAD/datasets/make_dataloader.py - Data loaders → Distributed sampler: uses distributed sampler
CVPR26_UAD/datasets/make_dataloader.py - Training entry → ReID model: builds model
CVPR26_UAD/train.py - ReID model → Vision Transformer: constructs backbone
CVPR26_UAD/model/make_model.py - Training entry → Identity and metric losses: builds losses
CVPR26_UAD/train.py - Training entry → Optimizer: builds optimizer
CVPR26_UAD/train.py - Training entry → Learning-rate schedule: creates schedule
CVPR26_UAD/train.py - Training entry → Training processor: starts training loop
CVPR26_UAD/train.py - Training processor → ProCA alignment: computes alignment
CVPR26_UAD/processor/processor.py - Training processor → GPD prototypes: updates and applies prototypes
CVPR26_UAD/processor/processor.py - Training processor → Retrieval metrics: evaluates retrieval
CVPR26_UAD/processor/processor.py - Retrieval metrics → Distance reranking: reranks distances
CVPR26_UAD/utils/metrics.py