Header background image

ORCA: Object Recognition and Comprehension for Archiving Marine Species

Yuk-Kwan Wong1, Haizin Liang1, Zeyu Ma2, Yiwei Chen1, Ziqiang Zheng1*,
Rinaldi Gotama3, Pascal Sebastian3, Lauren D. Sparks3, Sai-Kit Yeung1
1Hong Kong University of Science and Technology 2University of Electronic Science and Technology of China
3Indo Ocean Foundation
The IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) 2026
Oral Presentation
* Corresponding author: [email protected]
Paper Supplementary arXiv logo arXiv Hugging Face logo Dataset WACV Daily

Overview

Orca Overview
Figure 1. ORCA overview.

ORCA aims to enhance marine visual understanding for machine learning models. It aligns domain-specific challenges with core computer vision tasks, namely object detection, image grounding, and image captioning. The dataset includes 14,647 images covering 478 marine species (670 common names), annotated with 42,217 bounding boxes and 22,321 expert-verified positive captions, establishing a comprehensive benchmark for advancing computer vision in marine research.

Statistic

#. of Images #. of Species #. of Boxes #. of Captions (Refined / Positive / Negative)
14,647 478 42,217 12,873 / 9,948 / 12,431
Table 1. Summary statistics of the ORCA dataset. Captions are categorized as Refined (expert-refined), Positive (correct VLM-generated), or Negative (incorrect VLM-generated).
Orca comparison chart
Figure 2. ORCA offers a balance and sufficient amount of visual and textual annotations, compared to other general and marine datasets.
Dataset Image Count Visual
Annotation 		Linguistic
Annotation 		Category
Count 		Taxonomy
Supported
Marine Domain Datasets
DUO 7,782 BBOX - 4
SUIM 1,525 Mask - 8
MAS3K 3,103 Mask - 37
UIIS 4,628 Mask - 7
SEAMPD21 28,328 BBOX - 130
Wildfish 54,459 Category - 1,000
FishNet 94,532 BBOX - 17,357
Wildfish++ 2,348 Category Image-Level 2,348
General-purpose Datasets
Redcaps 12,011,121 - Image-Level -
Pascal Sentences 1,000 Category Image-Level 20
SBU Captions 1,000,000 - Image-Level -
iNat2017 859,000 BBOX - 5,089
Ours
ORCA (Ours) 14,645 BBOX Instance-Level 670
Table 2. Statistic comparison of ORCA with other general and domain-specific datasets. Notably, ORCA provides a comprehensive marine domain dataset with detailed instance-level annotations.
Orca caption chart
Figure 3. Caption tokens length comparison between general and marine datasets. ORCA features longer and dense caption annotations, compared to others.

Citation

                        
@InProceedings{Wong_2026_WACV,
    author    = {Wong, Yuk-Kwan and Liang, Haixin and Ma, Zeyu and Chen, Yiwei and Zheng, Ziqiang and Gotama, Rinaldi and Sebastian, Pascal and Sparks, Lauren D. and Yeung, Sai-Kit},
    title     = {ORCA: Object Recognition and Comprehension for Archiving Marine Species},
    booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
    month     = {March},
    year      = {2026},
    pages     = {1597-1609}
}