File size: 2,075 Bytes
286fbb6
 
fdfb620
 
 
 
286fbb6
fdfb620
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
---
license: cc-by-4.0
tags:
- ocean
- benthic
- object-detection
---

# FathomNet2023 Baseline Model

## Model Details
- Trained by researchers at [Monterey Bay Aquarium Research Institute](https://www.mbari.org/) (MBARI) as a baseline for the [FathomNet2023 Competition](https://www.kaggle.com/competitions/fathomnet-out-of-sample-detection/overview) presented with the [Fine Grained Visual Categorization workshop](https://sites.google.com/view/fgvc10/home) at CVPR 2023.
- [Ultralytics YOLOv8.0.117](https://github.com/ultralytics/ultralytics/pull/3145)
- Object detection
- Fine tuned yolov8m to detect 290 fine grained taxonmic categories of benthic animals in the Greater Monterey Bay Area off the coast of Central California.

## Intended Use
- Make detections on images collect on the sea floor in the Monterey Bay Area.

## Factors
- Distribution shifts related to sampling platform, camera parameters, illumination, and deployment environment are expected to impact model performance.
- Evaluation was performed on an IID subset of available training data.
- Data to test out of distribution performance can be found on the [competition Kaggle page](https://www.kaggle.com/competitions/fathomnet-out-of-sample-detection/overview).

## Metrics
- [Precision-Recall curve](https://huggingface.co/FathomNet/MBARI-midwater-supercategory-detector/blob/main/plots/PR_curve.png) and [per class accuracy]((https://huggingface.co/FathomNet/MBARI-midwater-supercategory-detector/blob/main/plots/confusion_matrix.png)) were evaluated at test time. 
- [email protected] = 0.33515
- Performance is quite variable depending on the target organism even when testing on in-distribution data.

## Training and Evaluation Data
- Training data is the [FathomNet2023 competition split](https://www.kaggle.com/competitions/fathomnet-out-of-sample-detection/overview) and internal MBARI data
- Class labels have a [long tail and localizations occur throughout the frame]().

## Deployment
In an environment running YOLOv8:

```
python classify/predict.py --weights best.pt --data data/images/
```