ICML 2026 · Datasets Track Detection · Panoptic · Tracking 10,101 high-resolution frames

MarPOT

A Large-Scale Maritime Panoptic Obstacle Tracking Benchmark

A maritime perception benchmark that unifies detection, panoptic segmentation, and multi-object tracking under temporally consistent instance identifiers.

Download Dataset Read the Paper

10,101

High-res images

52.7k

Bounding boxes

131.3k

Instance masks

Sequences

Orders of scale

MarPOT overview: scale variation, conditions, and unified annotations

What is MarPOT

Unified annotations across detection, segmentation, and tracking

MarPOT is built from six months of real maritime operations across harbor anchorages, coastal channels, and offshore routes.

🌊

Diverse Scenarios

Harbor anchorages, coastal channels, and offshore routes under varying weather, lighting, and traffic conditions.

🎯

Multi-Task Annotations

Bounding boxes, panoptic instance masks, and temporally consistent IDs — all aligned within the same frame.

⚓

Realistic Challenges

Occlusion, specular reflection, severe backlighting, haze, and extreme scale variations baked into the data.

Why a new benchmark

Existing maritime datasets isolate tasks. MarPOT unifies them.

Detection benchmarks provide only boxes. Segmentation benchmarks provide pixel labels but no temporal IDs. Tracking benchmarks lack dense instance masks.

No public maritime dataset previously offered temporally consistent panoptic identifiers.

See full comparison →

Object density across MarPOT sequences — Object density across 13 sequences exhibits the natural bimodal distribution of maritime operations.

Dataset Statistics

At a glance

10,101

High-resolution images

2800×2800 or 2800×1600, 15 fps source footage

52.7k

Bounding boxes

48,784 ships + 3,916 obstacles

131.3k

Polygon masks

52,700 things + 78,550 stuff; 13.0 per frame

Sequences

13 sequences across 6 months of operations

Orders of scale

Relative area from 10⁻⁴ to 10⁻¹

5.22

Avg. instances / frame

9.9 in port, 3.4 in open water

Key Features

Six properties that set MarPOT apart

Extreme Scale Variation

Object areas span four orders of magnitude. Distant vessels and nearby obstacles coexist in the same frame.

Complex Visual Conditions

Specular reflection, severe backlighting, fog, haze, and dense shoreline clutter.

Unified Annotation Framework

Occlusion / truncation attributes, polygonal instance masks, and temporal IDs share a single coordinate system per frame.

Temporal Consistency

Every dynamic instance keeps the same identifier across an entire sequence.

Geographic Diversity

Multiple regions and traffic patterns: dense harbors, transit channels, and sparse open water.

Comprehensive Baselines

22 semantic segmentation architectures, 2 panoptic frameworks across 6 backbones, and 7 multi-object trackers.

Sample Browser

Drag the slider to flip between RGB and panoptic mask

A curated subset of MarPOT frames covering port operations, coastal channels, offshore traffic, and adverse conditions.

RGB Mask

⇔

Harbor Anchorage at Sunrise camera1_0

13 anchored ships under low-angle sun with shoreline clutter and partial occlusions.

PortOcclusion

RGB Mask

⇔

Backlit Anchorage camera1_4198

Backlit fishing vessels moored near the shoreline under strong surface glare.

PortBacklitReflection

RGB Mask

⇔

Distant Return Wake 003614

A small vessel returning with a faint wake under uniform gray sky and sea.

CoastalSmall target

Browse all samples →

Download

Get the data

The full release ships in a single archive with images, panoptic masks, COCO-format JSONs, and the evaluation toolkit.

Full Dataset

Approx. 25 GB

10,101 images (2800×2800 / 2800×1600)
52,700 bounding-box annotations
131,250 polygonal masks
Temporal IDs across 13 sequences
Occlusion / truncation attributes

Download (cloud drive)

Mini Subset

Approx. 800 MB

~500 representative frames
All annotation types included
Evaluation scripts + quick-start docs

Download mini-set

Eval Toolkit

Code repository

Reference metrics: mIoU, PQ, HOTA, IDF1
Multi-scale mIoU + waterline accuracy
Submission templates for the leaderboard

GitHub repository

Terms of Use

MarPOT is released for academic research only. Commercial use requires written permission.

Citation

Cite MarPOT

If you use MarPOT in your work, please cite:

@inproceedings{marpot2026,
  title     = {MarPOT: A Large-Scale Maritime Panoptic Obstacle Tracking Benchmark},
  author    = {Anonymous Authors},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  year      = {2026}
}