Computer vision task
Image segmentation datasets and models
Pixel-precise polygon datasets and segmentation models.
Image segmentation labels an object at the pixel level rather than with a box: each instance is outlined by a polygon (or mask) that traces its exact shape. That precision matters when objects overlap, have irregular outlines, or when you need area and boundary measurements - defect maps, medical imaging, agriculture, and anything where a rectangle is too coarse.
Segmentation annotations export to the COCO segmentation, YOLO-seg, and mask formats. Pictograph's polygon and SAM 3 auto-polygon tools produce clean, hole-aware masks in a few clicks, and every dataset below can be forked and re-annotated in the browser.
Pictograph trains instance segmentation with RF-DETR and semantic segmentation with a PyTorch pipeline, exporting ONNX you can run anywhere. The datasets below carry polygon annotations that are ready to train on.
Datasets 6
Trained models 1
Segmentation FAQ
What is an image segmentation dataset?
An image segmentation dataset labels each object at the pixel level with a polygon or mask that traces its exact outline, rather than a bounding box. A model trained on it predicts those masks on new images, which is essential when you need precise shape, area, or boundary information.
What is the difference between instance and semantic segmentation?
Semantic segmentation assigns every pixel a class but does not separate individual objects, so two touching cars become one "car" region. Instance segmentation keeps each object distinct, giving one mask per car. Pictograph's polygon annotations support both, and it trains RF-DETR for instance masks and a PyTorch pipeline for semantic masks.
How do I create polygon masks quickly?
Use the SAM 3 auto-polygon tool: click a point, drag a box, or type a class name, and it returns a hole-aware polygon that you accept or refine. It removes most of the manual point-by-point tracing, and the result exports directly to COCO segmentation or YOLO-seg.
How do I train a segmentation model on Pictograph?
Fork or upload a dataset with polygon annotations, then start a training run and pick RF-DETR segmentation (instance) or the PyTorch semantic pipeline. Pictograph handles the export, GPU training, and ONNX packaging, and serves the model through the API or as a file.
Start building today
$5/mo free compute. No payment method required.