This repository contains a fine-tuned RF-DETR object-detection model for identifying derelict crab pots in side-scan sonar imagery. The model was trained on the PINGEcosystem/sss-crab-pot-detection-ds dataset, which contains annotated sonar imagery collected in Delaware's Inland Bays and Delaware Bay.
The model is part of the GhostVision effort to support scalable detection and mapping of derelict fishing gear from acoustic imagery.
📜 Publication
Bodine, C. S., Baxevani, K., Abbasi, N., Wierzbicki, J., Christoph, O., Hughes, C., Bagoren, O., Hines, O., Greco, J., & Trembanis, A. (2026). GhostVision: Democratizing Derelict Gear Detection Using Low-Cost Sonar and Artificial Intelligence. Journal of Marine Science and Engineering, 14(10), 951. https://doi.org/10.3390/jmse14100951
This model was trained from a dataset that originally contained both Crab-Pot and Maybe-Crab-Pot labels. During preprocessing, ambiguous Maybe-Pot examples were omitted so the exported detector predicts a single foreground class:
Crab-Pot
📦 Files
model.safetensors - serialized model weights
weights.onnx - ONNX export for portable inference
class_names.txt - class label mapping used for export
Data collected from northern Rehoboth Bay and Indian River Bay, Delaware
Export metadata highlights:
Model variant: rfdetr-small
Input resolution: 512
Exported label space: Crab-Pot
Dataset endpoint version: rx5YMJ3d3GZMFobFJf3Y/1
If you are looking for the source annotations, data splits, and schema details, use the dataset card above.
🎯 Intended Use
This model is intended for:
automated detection of derelict crab pots in side-scan sonar imagery
research workflows for marine debris mapping
benchmarking sonar object-detection pipelines
downstream human-in-the-loop review and prioritization
This model is not intended to replace field validation or expert review in operational removal workflows.
🧪 Inference Notes
The repository includes both native weights and an ONNX export to support different deployment targets.
The exported label space is:
Crab-Pot
Predictions should be interpreted in the context of sonar-specific variability such as substrate texture, tow geometry, acoustic shadowing, and target burial.
🚀 Usage
📡 Use with GhostVision
For end-to-end processing of side-scan sonar data, use this model through GhostVision, which handles sonar preprocessing, moving-window tiling, inference orchestration, optional tracking, and georeferencing.
Typical GhostVision workflow:
Install and launch GhostVision.
Open your sonar recording or batch folder.
Select the RF-DETR model in the model dropdown.
Run detection and review the exported detections, shapefiles, and waypoint products.
If you are working from the sibling GhostVision repository in this workspace, the relevant inference flow is implemented around ghostvision.detect, ghostvision.main_crabDetect, and ghostvision.class_crabObj_rf.
💻 Use without GhostVision
If you only want to run the detector itself, use the exported weights.onnx file directly with ONNX Runtime.
1from huggingface_hub import hf_hub_download
2import numpy as np
3from PIL import Image
4import onnxruntime as ort
56repo_id ="PINGEcosystem/gv-rf-detr"7model_path = hf_hub_download(repo_id=repo_id, filename="weights.onnx")89session = ort.InferenceSession(model_path, providers=["CPUExecutionProvider"])10input_name = session.get_inputs()[0].name
1112image = Image.open("path/to/sonar-image.jpg").convert("RGB").resize((512,512))13image_np = np.asarray(image, dtype=np.float32)/255.014image_np = np.transpose(image_np,(2,0,1))[None,...]1516raw_outputs = session.run(None,{input_name: image_np})17output_map ={18 output_meta.name: output_value
19for output_meta, output_value inzip(session.get_outputs(), raw_outputs)20}2122for name, value in output_map.items():23print(name, value.shape)
This direct ONNX path is useful when you want to embed the detector in another application or build your own post-processing pipeline. The exact output tensors depend on the export format, so inspect the returned names and shapes before writing box filtering and visualization code.
⚠️ Limitations
Performance may degrade on sonar systems, substrates, or regions that differ from the Delaware training domain.
Small, partially buried, or weak-return targets may be missed.
Sonar artifacts and hard-bottom features may produce false positives.
The model predicts only the retained foreground class used during export.
Outputs should be reviewed by domain experts before management or removal decisions are made.
