README.md

---
license: apache-2.0
language:
- en
pipeline_tag: image-segmentation
tags:
- pytorch
- rgb-d
- rail-surface-defect-detection
- sam2
- glide-net
---

# GLIDE-Net

Full PyTorch checkpoint for **Global-Local Interaction and Detail Enhancement
Network for RGB-D Rail Surface Defect Detection**.

Code and installation instructions:
[Jerry050512/GLIDE-Net](https://github.com/Jerry050512/GLIDE-Net).

## Model

GLIDE-Net uses a Transformer–CNN hybrid encoder with a frozen SAM2 Hiera-L
backbone and trainable adapters, Cross-Modal Fusion (CMF), and a Multi-Scale
Detail Enhancement and Fusion (MSDEF) decoder.

`glide_net_6440647.pth` contains the complete model state, including the frozen
backbone and batch normalization buffers. It requires no separate SAM2
checkpoint for inference. File size: 968,608,466 bytes.
This is a custom PyTorch model; use the implementation in the linked GitHub
repository to load it.

## Download and inference

```bash
pip install -U huggingface_hub
hf download jerry050512/GLIDE-Net glide_net_6440647.pth --local-dir checkpoints

git clone https://github.com/Jerry050512/GLIDE-Net.git
cd GLIDE-Net
# Install dependencies as described in the repository README.
python test.py \
    --data_path /path/to/NEU-RSDDS-AUG \
    --checkpoint /path/to/checkpoints/glide_net_6440647.pth \
    --device cpu --cpu_threads 4 --num_workers 0 \
    --save_path ./results/pred

python eval.py \
    --pred_path ./results/pred \
    --gt_path /path/to/NEU-RSDDS-AUG/GT_test
```

The dataset must contain aligned RGB BMP, depth TIFF and GT PNG files in
`Image_test/`, `Depth_test/` and `GT_test/`. Use the original dataset inputs,
rather than JPEG/PNG visualization copies. Input resolution is 352×352.
CPU inference is supported; GPU inference is faster.

## Results

The original experiment's saved predictions on the 362-image NEU RSDDS-AUG
test set yield:

| S-measure | MAE | max-E | max-F |
| --- | --- | --- | --- |
| 0.8718 | 0.0501 | 0.9475 | 0.9129 |

max-F uses the maximum of the averaged F-measure curve with beta=0.3;
max-E uses the maximum of the averaged E-measure curve.
The original experiment used the test split for checkpoint selection.

The checkpoint was selected at epoch 27 and migrated from experiment commit
`64406472f9c90dbf30f4ba44115c1e3a62167699` by renaming all 991 state tensors
without changing their values. Three sampled old/release CPU model outputs
matched exactly at all decoder scales. An 85-image CPU check matched historical
prediction metrics at four decimal places; some pixels differed by one uint8
gray level. This is a subset verification, not a full CPU rerun of 362 images.

## Intended use and limitations

The model is intended for research on RGB-D rail surface defect segmentation.
Performance on other cameras, depth sensors or datasets has not been established.
It has not been validated for autonomous safety-critical rail inspection.

## License and attribution

Released under Apache-2.0; see [LICENSE](LICENSE).
The included frozen backbone derives from
[Meta's SAM2](https://github.com/facebookresearch/sam2), also Apache-2.0.
Copyright (c) Meta Platforms, Inc. and affiliates. All rights reserved.
Dataset rights remain with the original dataset authors.

The checksum of the uploaded checkpoint is provided in `SHA256SUMS.txt`.
Formal citation details will be added after publication.