// the find
daberpro/lantern-image-loader
A C++ library for high-performance, multithreaded image loading and processing. It features an efficient producer-consumer architecture, automatic image resizing, and seamless integration with ArrayFire for machine learning data pipelines.
A single-header C++ library that loads images off a background thread into a fixed-size in-memory cache, with optional conversion to ArrayFire arrays for ML pipelines. It's aimed at someone building a small training loop who wants image I/O off the main thread without pulling in a full data-loading framework.
The producer-consumer queue with a blocking Get() is a reasonable pattern for overlapping I/O with compute, and template-parameterizing width/height/channel count avoids runtime branching in the hot resize path. Folder-based auto-labeling (parent directory as class name) plus CSV label loading covers the two most common dataset layouts without extra config. Header-only with stb_image/stb_image_resize2 keeps the dependency footprint small if you don't need the ArrayFire path.
TOTAL_IMAGES is a compile-time template parameter, so the cache size is baked into the type — you can't size it to an arbitrary dataset without recompiling, and it's unclear from the README what happens when a dataset exceeds that count. There's a hand-rolled lantern::utility::Vector instead of std::vector with no explanation of what it does differently, which is a red flag for correctness and maintenance. No license file, no tests, no CI, and 0 stars/forks on a repo last pushed in August — this is unproven and there's no evidence anyone but the author has run it. The API also silently assumes single-producer/single-consumer usage (switching datasets requires manually calling Stop() then Run() again) with no documented behavior for concurrent access or error handling on corrupt/missing images.