icevision

Why we included this project

IceVision does not tie you to a single training library. It puts COCO, VOC, and other annotation formats behind one consistent Python API, with bundled parsers and transforms, and hands the pipeline to either Fastai or PyTorch-Lightning, so switching frameworks does not mean rewriting your data-loading code. Pretrained weights from several model collections are reachable through the same interface, which makes comparing architectures surprisingly easy. The dataset inspection dashboard and automated cleaning helpers catch bad labels and file problems before training starts, a real time-saver on messy real-world data. It also supports multi-task training, so detection, segmentation, and classification heads can run together, which suits research teams and anyone building custom detection pipelines who would rather not assemble all the pieces by hand.