rifqijuli/soda_can_asta

BlueROV2 Underwater Object Detection — Test Set

A first-person underwater video test set captured with a BlueROV2 Heavy ROV, annotated in YOLO format and re-labelled to match four established underwater object detection datasets. Intended for evaluating pre-trained models from those datasets on real-world ROV footage without retraining.

---

Dataset Summary

The footage was recorded across 15 distinct motion sequences (forward, yaw, ascend, descend, diagonal) in an underwater environment. Frames were extracted and annotated using CVAT with two object classes: Can and Bag.

To allow direct evaluation against models trained on existing public datasets, the annotations were re-mapped to match the class schemes of four target datasets. Each ZIP in this repository contains the same images paired with a different annotation scheme.

File	Target Dataset	Classes	Annotation
cou_yolo.zip	COU Dataset	24	Soda Can, Plastic Bag, …
trashcan_yolo.zip	TrashCan 1.0	22	trash_can, trash_bag, …
uno_yolo.zip	UNO Dataset	38	metal_can, plastic_bag, …
walia_yolo.zip	Walia et al.	3	Trash, Rov, Bio

---

Repository Structure

Each ZIP follows the standard YOLO folder structure and can be used directly with Ultralytics:

<dataset>_yolo.zip
├── images/
│   └── test/
│       ├── c_d_frame_0001.jpg   ← center_descend sequence
│       ├── c_d_frame_0002.jpg
│       ├── c_f_frame_0001.jpg   ← center_forward sequence
│       └── ...
├── labels/
│   └── test/
│       ├── c_d_frame_0001.txt   ← YOLO format: class cx cy w h
│       ├── c_d_frame_0002.txt
│       └── ...
└── data.yaml                    ← class names matching the target dataset

All four ZIPs share identical images. Only labels/ and data.yaml differ between them.

---

data.yaml

Each ZIP includes a ready-to-use data.yaml. Example for COU:

path: /your/path/to/cou_yolo
train: images/test
val: images/test
test: images/test
nc: 24
names:
  0: Unknown Instance
  1: Scissors
  2: Plastic Cup
  ...
  6: Soda Can
  8: Plastic Bag
  ...

train and val are set to images/test intentionally — this dataset is a test-only set. Ultralytics requires all three splits to be defined.

---

Sequences

Frames are prefixed by sequence name for traceability:

Prefix	Sequence	Motion type
c_d_	center_descend	Descend
c_f_	center_forward	Forward
c_y_	center_yaw	Yaw
e_f_	east_forward	Forward
n_f_	north_forward	Forward
n_y_	north_yaw	Yaw
ne_f_	northeast_forward	Diagonal forward
nw_f_	northwest_forward	Diagonal forward
s_f_	south_forward	Forward
s_y_	south_yaw	Yaw
se_f_	southeast_forward	Diagonal forward
se_ya_	southeast_yawascend	Yaw + ascend
sw_f_	southwest_forward	Diagonal forward
sw_ya_	southwest_yawascend	Yaw + ascend
w_f_	west_forward	Forward

---

Usage

Evaluate a pre-trained model (Ultralytics)

from ultralytics import YOLO

model = YOLO("yolo11s_cou.pt")  # model trained on COU dataset
metrics = model.val(
    data="cou_yolo/data.yaml",
    split="test",
    conf=0.5,
    iou=0.5,
)
print(metrics.box.map50)

CLI

yolo val model=yolo11s_cou.pt data=cou_yolo/data.yaml split=test conf=0.5 iou=0.5

---

Annotation Details

• Format: YOLO (normalised class cx cy w h, one .txt per frame)
• Image resolution: 1920 × 1080
• Original annotation tool: CVAT (MOT export → converted to YOLO)
• Source classes: Can (MOT ID 1), Bag (MOT ID 2)
• Conversion: MOT absolute bbox → YOLO normalised; class IDs remapped per target dataset

---

License

Images and annotations are released under CC BY 4.0.
If you use this dataset, please cite the corresponding target dataset alongside this one.

---

Citation

@dataset{bluerov2_test_set_2025,
  author    = {Rifqi, [Your Last Name]},
  title     = {BlueROV2 Underwater Object Detection Test Set},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/[your-username]/bluerov2-test-set}
}

---

Acknowledgements

Annotation remapping to target dataset formats was performed using a custom browser-based tool built for this thesis. Target dataset class schemes are credited to their respective original authors.