Skip to content

About

AnnotateIt practice datasets, source images and model comparison evidence. CC BY 4.0 with attribution.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Coastal Scene: 100 images for annotation practice

100 image entries from one coastal recording, each 1820 x 1136 pixels. There are 92 unique image files by SHA-256: eight entries repeat a frame exactly. The original 100-entry selection is retained to match the video. Adjacent frames are correlated. The image owner authorized redistribution with credit and a link to AnnotateIt on 2026-09-30.

Attribution and license

Coastal Scene by AnnotateIt — https://annotateit.ai/datasets/coastal-scene/ — CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). Indicate any changes you make.

Images and AnnotateIt's accompanying dataset documentation are licensed under CC BY 4.0. This is not a license for model weights, third-party software, or the AnnotateIt application.

Coastal source frame

Download version 1

Files

  • coastal-scene-starter-v1.zip: AnnotateIt project, 100 images, no labels or annotations.
  • coastal-scene-images-v1.zip: the same unmodified images with ordinary .jpg filenames.
  • coastal-scene-comparison-evidence-v1.zip: three source frames, saved model polygon coordinates, original preview SVGs, rendered comparison panels and settings from the recorded experiment. This is a diagnostic evidence pack, not an importable labeled project or a COCO export.

Import the starter through Create from exported project. Use Auto-label to discover supported classes, or add car and boat for the two-class AI workflow. The Windows manual guide instead uses building, boat, car, bridge and seawall.

What this release does not contain

The video records 757 RF-DETR draft annotations across 100 frames. This release does not contain that complete batch export. The archived comparison has three frames per model; its local preview SVGs do not retain class assignments or confidence scores. Coordinates are provided without inventing those missing fields. No annotations are certified as human-reviewed ground truth.

The frames are highly correlated. Do not use a random split of these frames to claim independent model evaluation. Use separate recordings or locations for held-out evaluation.

Settings

ECSeg-M, ECSeg-X and RF-DETR-Seg-2XL: preview confidence 0.40. Additional RF-DETR preview: 0.25. The recorded batch applied an additional 0.50 output filter. ChatGPT: GPT-6-Astra, car/boat labels, 32 draft polygons across three frames. Hardware, historical app/runtime versions and weight hashes were not captured together. The material supports visual inspection of this run, not a reproducible accuracy benchmark.

Read and watch

Checksums are in SHA256SUMS.txt. Version 1.0.0, released 2026-09-30.

See comparison-manifest.json for model-key definitions and the exact matching source-image filenames.

About

AnnotateIt practice datasets, source images and model comparison evidence. CC BY 4.0 with attribution.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors