COCO128 Example Datasets

One small dataset per task, ready to train

Small COCO subsets, one for every task that LightlyTrain supports. They all come from the first 128 images of COCO train2017 and val2017.

Each task folder holds a config.yaml. Pass it straight to data=:

import lightly_train

lightly_train.train_object_detection(
    out="out/my_experiment",
    model="ltdetrv2-s-coco",
    data="object_detection/config.yaml",
)

Datasets

Task	Folder	Images used	Labels
Object detection	object_detection/	128 train, 128 val	YOLO boxes, class ids 0-79
Instance segmentation	instance_segmentation/	100 train, 100 val	YOLO-seg polygons, 80 thing classes
Panoptic segmentation	panoptic_segmentation/	100 train, 100 val	COCO panoptic JSON and mask PNGs
Semantic segmentation	semantic_segmentation/	100 train, 100 val	Class-index PNG masks, 130 classes
Pretraining, distillation	pretrain_distill/	128 train	none

The images/ pool always holds 128 train and 128 val images. Provenance explains why three of the tasks label only 100 of them.

Requirements

lightly-train 0.16.3 or later. Older versions resolve a relative data= path against the working directory, not against the location of the config file. This breaks the ../images/... paths that the mask-based configs use.

Download

A labeled task needs two archives: the shared image pool, and the labels for that task. Unzip both into the same directory.

# The image pool. Every task needs it.
wget https://github.com/lightly-ai/coco128_yolo/releases/download/v0.0.2/images.zip && unzip -q images.zip

# The labels. Take the line for your task.
wget https://github.com/lightly-ai/coco128_yolo/releases/download/v0.0.2/object_detection.zip && unzip -q object_detection.zip
wget https://github.com/lightly-ai/coco128_yolo/releases/download/v0.0.2/instance_segmentation.zip && unzip -q instance_segmentation.zip
wget https://github.com/lightly-ai/coco128_yolo/releases/download/v0.0.2/panoptic_segmentation.zip && unzip -q panoptic_segmentation.zip
wget https://github.com/lightly-ai/coco128_yolo/releases/download/v0.0.2/semantic_segmentation.zip && unzip -q semantic_segmentation.zip

Pretraining and distillation use no labels. For those two, images.zip is the whole download. Point data at images/train2017.

Layout

The archives carry no wrapper directory. After you unzip, the task folder and images/ are siblings:

.
├── images/
│   ├── train2017/          128 images
│   └── val2017/            128 images
└── object_detection/
    ├── config.yaml
    ├── images -> ../images
    └── labels/

Each config points at images/ by a relative path, so this layout is what makes the paths resolve. Run your training script from this directory. Then reference the data as data="object_detection/config.yaml".

All five tasks read the same images/ pool. You download the image bytes one time, for any number of tasks.

The images symlink

In object_detection/ and instance_segmentation/, the images entry is a symlink to ../images. It is not a real directory. The YOLO loader in lightly-train takes the labels directory from the same base path as the images directory. The two must therefore be siblings inside one task folder.

Most versions of unzip keep the symlink. Some Windows zip tools write it out as a text file that holds the target path. If that happens, make the link by hand:

ln -s ../images object_detection/images         # macOS, Linux
mklink /D object_detection\images ..\images     # Windows

Do the same for instance_segmentation/images.

Provenance

• object_detection/, pretrain_distill/: the original first-128-image YOLO and unlabeled subsets. Only the folder location changed.
• panoptic_segmentation/: migrated from the original coco128_panoptic repo. See scripts/trim_panoptic.py.
• instance_segmentation/: YOLO-seg polygon labels, built from the COCO panoptic annotations in panoptic_segmentation/ (thing classes only). See scripts/build_instance_segmentation.py. The class ids are the COCO panoptic category ids, not the contiguous 0-79 ids that object_detection/ uses.
• semantic_segmentation/: single-channel class-index PNG masks, built from the same panoptic annotations (thing and stuff classes). See scripts/build_semantic_segmentation.py. Class id 0 means "unlabeled".

Why three tasks label only 100 images

The upstream coco128_panoptic repo listed 128 annotation entries per split, but shipped only 100 mask PNGs. The upload was incomplete. Both splits were trimmed to the 100 images that have a mask. instance_segmentation/ and semantic_segmentation/ come from panoptic_segmentation/, so all three cover the same 100 train and 100 val images.

The images/ pool keeps all 128 images per split, because object detection and pretraining need them. The mask-based loaders in lightly-train skip an image with no matching mask, and instance_segmentation/config.yaml sets skip_if_label_file_missing: true for the same reason. scripts/trim_panoptic.py now trims the panoptic annotation JSON only. It no longer deletes images.

Maintainers

See RELEASE.md for how to build the download assets and cut a release.

About Lightly

Lightly builds open-source tools for computer vision teams.

    

• LightlyTrain — pretraining, fine-tuning and distillation. The datasets here are its example data.
• LightlyStudio — curate, annotate and manage image datasets.
• LightlySSL — self-supervised learning framework.

Questions go to Discord. For commercial use, contact us.