Fine-tuned YOLOv8m-seg for underwater debris detection and instance segmentation, specifically designed for ROV (Remotely Operated Vehicle) camera footage — seabed and underwater surface debris, not floating surface trash.

Project: Life Under Water Author: Krishna Jaiswal Live Demo: Hugging Face Spaces

Results

Metric	Score
mAP50 — detection	65.6%
mAP50 — segmentation	65.0%
Precision	79%
Classes	16
Training images	6,008
Validation images	1,204
Epochs	75

Sample Predictions

Training Curves

Dataset

TrashCan 1.0 — Instance segmentation dataset of underwater trash captured by JAMSTEC deep-sea ROV cameras. Source: University of Minnesota Data Repository

16 Classes

Category	Classes
Trash	`trash_plastic`, `trash_metal`, `trash_fabric`, `trash_fishing_gear`, `trash_rubber`, `trash_wood`, `trash_paper`, `trash_etc`
Marine life	`animal_fish`, `animal_starfish`, `animal_shells`, `animal_crab`, `animal_eel`, `animal_etc`, `plant`
Equipment	`rov`

Class Distribution (training set)

Class	Instances	%
rov	2,653	44.2%
trash_etc	1,629	27.1%
trash_plastic	1,490	24.8%
trash_metal	901	15.0%
animal_fish	611	10.2%
plant	405	6.7%
animal_starfish	274	4.6%
trash_wood	271	4.5%
animal_eel	267	4.4%
animal_crab	247	4.1%
trash_fabric	247	4.1%
animal_etc	180	3.0%
animal_shells	171	2.8%
trash_paper	154	2.6%
trash_fishing_gear	127	2.1%
trash_rubber	113	1.9%

Note: Severe class imbalance (ROV 44%, trash_rubber 2%) — use conf=0.15 for better detection of rare trash classes.

Usage

python

from ultralytics import YOLO
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(
    repo_id="Krishna-Jaiswal/yolov8m-marine-trash",
    filename="yolov8m_marine_best.pt"
)

# Run inference
model   = YOLO(model_path)
results = model.predict("underwater_image.jpg", task="segment", conf=0.15)
results[0].show()

Training Configuration

Parameter	Value
Model	YOLOv8m-seg (pretrained COCO)
Optimizer	SGD
Learning rate	0.01 → 0.001 (cosine decay)
Epochs	75 (early stopping patience=20)
Batch size	16
Image size	640×640
Augmentation	Mosaic, Mixup=0.1, Copy-paste=0.1, HSV jitter, Flip
Platform	Kaggle T4 GPU

Limitations

Trained specifically on JAMSTEC ROV deep-sea footage — performance may vary on footage from different cameras or water conditions
Class imbalance: ROV (44%) dominates vs trash_rubber (2%) — model biased toward ROV
mAP (65.6%): Realistic for challenging underwater domain (low-light, murky conditions, occlusions)
Best performance on seabed debris, not floating surface trash
Per-class mAP varies: ROV ~90% vs rare trash classes ~20-40%

Real-world Applications

Underwater ROV-based ocean cleanup systems
Marine pollution monitoring and research
Automated seabed debris surveys
Aligned with UN SDG 14 — Life Below Water
Same problem domain as EU SeaClear 2.0 (€9M Horizon Europe project)

Citation

bibtex

@dataset{trashcan2020,
  title  = {TrashCan 1.0: An Instance-segmentation Labeled Dataset of Trash Observations},
  author = {Hong, Jungseok and Fulton, Michael and Sattar, Junaed},
  year   = {2020},
  url    = {https://conservancy.umn.edu/handle/11299/214865}
}