Lightly

What is it

Lightly is a computer-vision toolset from Switzerland designed for machine-learning teams handling data. It covers three aspects: data curation (picking the most valuable, representative samples), self-supervised pretraining (letting a model learn useful representations without large amounts of labels), and smart data selection at the edge (filtering, on-device, which data is worth collecting back for training).

What problem it solves

When training computer-vision models, the real bottleneck is often not lack of data but too much data, too much redundancy, and too-high labeling cost. Blindly labeling and training wastes budget and drags the model down with masses of redundant samples. Lightly helps teams pick the truly informative samples from huge raw image sets, reducing labeling volume, improving training efficiency, and using unlabeled data via self-supervised pretraining. It suits ML teams, data engineers, and researchers building or iterating computer-vision models, especially with large data volumes, limited labeling budgets, or needing to continuously collect data on edge devices. If you want to spend resources where they matter rather than indiscriminately labeling every image, this kind of data-curation tool is very valuable.