README.md

# MDF-Det and reproduced-baseline model weights

This directory is prepared for a future Hugging Face model repository. It is
kept outside the GitHub source-code clone and must not be added to the MDF-Det
Git repository.

## Contents

| Directory | Files | Purpose |
|---|---|---|
| `MDF-Det/` | Binary classifier, normalization, SA-TD regression model, regression normalization and SPGF-V4 | Complete revised MDF-Det inference |
| `Dual-CNN/` | Binary classifier, regression model and their normalization files | Original two-CNN baseline inference |
| `baselines/ClusterNet/` | `best.pt` | ClusterNet + FoveaNet reproduction |
| `baselines/HMRN/` | `best.pt` | HMRN reproduction |
| `baselines/HM-Net/` | `best.pt` | HM-Net reproduction |
| `baselines/CATLoss/` | `best.pt` | CATLoss reproduction |

TTE-KH is intentionally excluded. Only the selected best checkpoints are
included; latest checkpoints, optimizer backups, ablation weights, caches and
experimental results are omitted.

## Source-code mapping

After downloading, either update each command-line/configuration checkpoint
path or copy the files to the expected local locations:

- MDF-Det binary classifier and normalization correspond to
  `Models/BinaryClassification/saved_model_2.model` and
  `saved_image_norm_2.model`.
- MDF-Det SA-TD files correspond to `regression_spatial_attention.h5` and
  `regression_norm_params.npz`.
- MDF-Det SPGF corresponds to
  `models/semantic_prior_v4_six_aoi_excluded/best_semantic_prior_v4.h5`.
- Reproduced PyTorch baselines use the `checkpoint` paths in their JSON
  inference configurations under the public `baselines/` source directory.

## Integrity

`MANIFEST.csv` records the byte size and SHA-256 digest of every weight file.
Verify downloaded files before inference. The current bundle contains 13 model
artifacts and is approximately 419.63 MiB.

## Dataset

The WPAFB imagery and annotations are not part of this bundle. For research
dataset access, contact likangqiushi20@nudt.edu.cn.