|
--- |
|
license: mit |
|
tags: |
|
- FasterRCNN |
|
- PyTorch |
|
- Marine |
|
--- |
|
# Model Card for Handfish Detector |
|
|
|
<!-- Provide a quick summary of what the model is/does. --> |
|
|
|
A Faster RCNN object detector for handfish, a rare bony fish found in the waters around Tasmania, Australia. |
|
|
|
## Model Details |
|
|
|
### Model Description |
|
- **Developed by:** Heather Doig, Australian Centre for Robotics, University of Sydney |
|
- **Model type:** Faster RCNN object detector from Detectron2 with ResNet50 with FPN backbone |
|
- **License:** MIT |
|
- **Finetuned from model [optional]:** [detectron2://COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl](detectron2://COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl) |
|
|
|
## Uses |
|
|
|
For use on images captured by Autonomous Underwater Vehicles in the [IMOS](https://imos.org.au/) program. |
|
|
|
## Training Details |
|
|
|
### Training Data |
|
|
|
Model trained on 200 images with bounding box annotations of handfish from Squidle+ [squidle.org](squidle.org) as well as 2000 synthetic images of handfish in underwater scenes generated by Blender and Infinigen [infinigen.org](infinigen.org) |
|
|
|
### Training Procedure |
|
|
|
Trained with detectron2 with student-teacher model using exponential moving average to update the weights of the teacher. |
