fiftyone

Why we included this project

If your team builds image or video models, FiftyOne is where the messy day-to-day work happens. It pairs a Python API with an interactive app, letting you browse datasets and filter by labels or predictions. When a run looks wrong, you can drill into the exact slices where the model is failing instead of stitching together ad-hoc scripts to visualize detections or inspect embeddings. Before data ever reaches training, you get a single place to curate and label it, then evaluate how well your model handles it. Teams doing error analysis on detection or classification runs, or preparing clean datasets to hand off to their pipelines, tend to find it useful right away. It also sits on top of whatever model framework you already use, so adopting it doesn't mean changing your stack.