Model Card for model_for_rail_anomaly_detection

Model Summary

This repository contains a fine-tuned version of I-JEPA (Image-based Joint-Embedding Predictive Architecture) designed for visual anomaly detection in railway environments.

The model leverages self-supervised visual representation learning to capture the underlying visual structure of normal railway scenes. Its objective is to learn meaningful latent representations of railway tracks and their surrounding environments, enabling the identification of potentially anomalous situations, such as obstacles appearing on the tracks.

The intended application is railway infrastructure monitoring, particularly camera-based obstacle detection systems operating in real-world and potentially challenging conditions.

Model Details

Model Description

• Developed by: Ilyas DAHAOUI
• Model type: JEPA-based Visual World Model
• Base architecture: I-JEPA (Image-based Joint-Embedding Predictive Architecture)
• Model family: Vision Transformer (ViT)
• Learning paradigm: Self-supervised learning
• Fine-tuning approach: Domain adaptation on railway imagery
• Framework: PyTorch / Hugging Face Transformers
• Primary task: Visual anomaly detection
• Target domain: Railway infrastructure and track monitoring
• Input modality: RGB images
• Output: Latent visual representations

Intended Use

The model is designed to support the development of camera-based railway obstacle detection systems.

Its primary objective is to learn representations of normal railway environments without relying on large-scale manually annotated datasets.

Potential anomalous situations include:

• Objects or debris on railway tracks.
• Fallen trees or other infrastructure-related obstacles.
• Animals or humans present on the tracks.
• Unusual visual patterns deviating from normal railway scenes.

The model can be integrated into downstream anomaly detection pipelines using embedding comparison, feature-distance metrics, density estimation, or other anomaly scoring techniques.

Training Approach

The model is based on I-JEPA, a self-supervised architecture that learns visual representations by predicting the latent representations of target image regions from contextual image regions.

Unlike conventional supervised approaches, I-JEPA does not require explicit semantic labels to learn visual representations.

The model was fine-tuned on railway imagery to adapt its learned representations to the visual characteristics of railway environments.

The fine-tuning process aims to improve domain-specific representation learning, particularly the ability to capture the structural and contextual patterns associated with normal railway scenes.

Limitations

• The model produces latent representations rather than direct anomaly classifications.
• An additional anomaly scoring or decision mechanism is required to distinguish normal and anomalous observations.
• Performance may vary under significant illumination changes, adverse weather conditions, camera variations, motion blur, or previously unseen environments.
• The model should be evaluated on representative railway scenarios before deployment in operational environments.
• Fine-tuning on normal railway imagery does not guarantee the detection of every possible anomaly.

Intended Audience

• AI and computer vision researchers.
• Railway infrastructure monitoring engineers.
• Autonomous transportation researchers.
• Developers working on self-supervised anomaly detection and world models.

Usage

from transformers import AutoModel, AutoImageProcessor
from PIL import Image

model_id = "ilyass31/model_for_rail_anomaly_detection"

processor = AutoImageProcessor.from_pretrained(model_id)
model = AutoModel.from_pretrained(model_id)

image = Image.open("railway_image.jpg").convert("RGB")

inputs = processor(images=image, return_tensors="pt")

outputs = model(**inputs)

embeddings = outputs.last_hidden_state
print(embeddings.shape)

Ethical Considerations

This model is intended to support railway safety research and infrastructure monitoring. Its outputs should not be considered a standalone safety decision.

Any operational deployment should include additional validation, anomaly detection mechanisms, and appropriate safety procedures.

Citation

If you use this model in your research or projects, please cite this repository:

@misc{dahaoui_rail_ijepa,
  author = {Dahaoui, Ilyas},
  title = {I-JEPA Fine-Tuned for Railway Anomaly Detection},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/ilyass31/model_for_rail_anomaly_detection}
}