Marine litter poses a growing threat to the health of our planet. Our planet's oceans and lakes are choked with plastic and other debris, fed to them through equally polluted rivers and streams, carried in by runoff, or dropped off ships. The pollution of these water sources can devastate entire populations of aquatic, terrestrial, and plant life. While preventing debris from entering the water at all is the only sustainable solution to this problem, clean up efforts are crucial to solving the problem or at least slowing its spread.
To aid in the application of robots to cleanup efforts, we have compared four popular deep neural networks for object detection, trained to detect trash using data drawn from a dataset of underwater ROV dive videos which can be found here.
A video of our results is included below, along with tables comparing the performance and accuracy of all four networks, trained by fine tuning.
| Network | mAP | Avg. IOU | plastic | bio | rov |
|---|---|---|---|---|---|
| YOLOv2 | 47.9 | 54.7 | 82.3 | 9.5 | 52.1 |
| Tiny-YOLO | 31.6 | 49.8 | 70.3 | 4.2 | 20.5 |
| Faster-RCNN | 81.0 | 60.6 | 83.3 | 73.2 | 71.3 |
| SSD | 67.4 | 53.0 | 69.8 | 6.2 | 55.9 |
| Network | 1080 | TX2 | CPU |
|---|---|---|---|
| YOLOv2 | 74 | 6.2 | 0.11 |
| Tiny-YOLO | 205 | 20.5 | 0.52 |
| Faster-RCNN | 18.75 | 5.66 | 0.97 |
| SSD | 25.2 | 11.25 | 3.19 |
A newer addition to the work is the evaluation of transfer learning as a method to overcome the limited amount of data available for this problem. YOLOv2 was trained using fine tuning, transfer learning on the last four and last three layers only.
| Training Method | mAP | Avg. IOU | plastic | bio | rov |
|---|---|---|---|---|---|
| Fine Tuning | 47.9 | 54.7 | 82.3 | 9.5 | 52.1 |
| Last 4 Layers | 33.9 | 45.5 | 71.3 | 13.6 | 17.0 |
| Last 3 Layer | 39.5 | 34.1 | 74.6 | 19.9 | 23.9 |
More information on the specifics of each class and the construction of the data model can be found in the relevant papers on the Publications page.