Can I compare multiple models on the same dataset?

Absolutely. The skill encourages providing unique label field names for each model run, allowing you to compare predictions from different architectures side-by-side in the FiftyOne App.

Does this skill help with visualizing inference results?

Yes, the workflow automatically handles launching and closing the FiftyOne App, allowing you to visually inspect predictions and filtered views immediately after inference completes.

Do I need to import my dataset before running inference?

Yes, the skill includes a verification step that checks for the dataset's existence and will guide you to use the fiftyone-dataset-import skill if the data hasn't been loaded yet.

What models are supported by this skill?

It supports a wide variety of models from the FiftyOne Model Zoo, including YOLOv8, CLIP, SAM, DINOv2, Faster R-CNN, and Vision Transformers for tasks like detection, classification, and embeddings.

How does the skill handle missing Python dependencies?

If a model requires an external library like 'ultralytics' or 'segment-anything', the skill detects the missing package and provides the exact pip install command to resolve the issue.

FiftyOne Dataset Inference

Name: FiftyOne Dataset Inference
Author: voxel51

byvoxel51

•

Data Science & ML

Runs machine learning model inference, including detection, classification, and segmentation, directly on FiftyOne datasets.

About

The FiftyOne Dataset Inference skill empowers developers to apply high-performance computer vision models from the FiftyOne Model Zoo to their datasets seamlessly. It supports a wide range of tasks such as object detection, image classification, instance segmentation, and embedding generation using industry-standard architectures like YOLOv8, CLIP, SAM, and DINOv2. By automating the end-to-end workflow—from dataset verification and dependency management to launching the FiftyOne App and executing operators—this skill streamlines the process of generating and visualizing model predictions, making it essential for data curation, model evaluation, and ground-truth comparison.

Key Features

Execute inference using popular models like YOLOv8, CLIP, Segment Anything (SAM), and ResNet.
Support for object detection, image classification, segmentation, and visual embedding generation.
Automated FiftyOne App management for real-time visualization of model predictions.
Flexible field mapping to prevent overwriting ground truth and enable multi-model comparisons.
3 GitHub stars
Built-in dependency detection with automated installation commands for missing ML packages.

Use Cases

Comparing multiple model variants (e.g., YOLOv8n vs YOLOv8x) against ground truth for performance benchmarking.
Running automated object detection on new datasets to generate initial bounding box labels.
Generating CLIP or DINOv2 embeddings for large-scale visual similarity search and data pruning.

About

Key Features

Execute inference using popular models like YOLOv8, CLIP, Segment Anything (SAM), and ResNet.
Support for object detection, image classification, segmentation, and visual embedding generation.
Automated FiftyOne App management for real-time visualization of model predictions.
Flexible field mapping to prevent overwriting ground truth and enable multi-model comparisons.
3 GitHub stars
Built-in dependency detection with automated installation commands for missing ML packages.

Use Cases

Comparing multiple model variants (e.g., YOLOv8n vs YOLOv8x) against ground truth for performance benchmarking.
Running automated object detection on new datasets to generate initial bounding box labels.
Generating CLIP or DINOv2 embeddings for large-scale visual similarity search and data pruning.