Building an autonomous vessel in marrine envrionment have so many challenges. One of the challenges is providing the vessel with the vision capabilities so it can analyze the marrine traffic and make different decisions i:e navigation, collision detection etc. This project is initiated to handle this problem. It focuses on following
1. Detect and classify maritime vessels using a vision based system
2. Determine the hull color
3. Detect name of ship
The dataset was collected by scraping images from internet. It contains 11 classes.
| Classes |
|---|
| Tanker |
| Tug |
| Fishing Vessel |
| Container |
| Passenger Ship |
| Sailing Vessel |
| Military Vessel |
| Supply Vessel |
| Power Boat |
| Jet Ski |
| Model | Inference time | mAP | Issues |
|---|---|---|---|
| Yolov3 | 40 ms | 68.7 | high false positive rate |
| Yolov3-Tiny | 10 ms | 62 | Produces extra detections (high false Negatives) |
| RetinaNet | 100 ms | 84.3 | Slow and need high GPU memory |
Open terminal and browse into oceans11 repo
pip3 install -r requirements.txt
cd <path-to-repo>/darknet
make
To use CUDA follow original instructions here https://pjreddie.com/darknet/install/
In yolo_inference.py change the path of weights and input_video to run inference on a video