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
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

                        
@misc{wong2025orcaobjectrecognitioncomprehension,
      title={ORCA: Object Recognition and Comprehension for Archiving Marine Species}, 
      author={Yuk-Kwan Wong and Haixin Liang and Zeyu Ma and Yiwei Chen and Ziqiang Zheng and Rinaldi Gotama and Pascal Sebastian and Lauren D. Sparks and Sai-Kit Yeung},
      year={2025},
      eprint={2512.21150},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2512.21150}, 
}