// the find
alibaba/EasyCV
An all-in-one toolkit for computer vision
EasyCV is Alibaba's all-in-one PyTorch CV toolkit covering self-supervised learning (SimCLR, MoCo v2, DINO, MAE), image classification (including the full timm model catalog), detection (YOLOX-PAI, DETR/DAB-DETR/DINO), segmentation (Mask2Former), and even 3D detection (BEVFormer) under one config-driven framework. It's for teams that want one toolkit across several CV tasks instead of stitching together separate repos per architecture family, especially if you're already used to OpenMMLab-style configs.
Genuinely broad model zoo in a single framework rather than five separate repos — SSL, classification, detection, segmentation, and 3D detection all share the same dataset/model/hook abstraction. It wraps the entire timm catalog instead of reimplementing backbones, saving real porting work. Production features go beyond research code: DALI for data loading, JIT export, PAI-Blade inference optimization, and a batch-prediction API aimed at actually serving models, not just training them.
The public changelog stops at v0.11.0 (May 2023) despite a more recent push date, so recent activity looks like dependency bumps and fixes rather than active feature work — budget for maintenance mode, not momentum. Most of the deeper technical writeups are Zhihu posts in Chinese with no English equivalent, so you hit a documentation wall fast once you're past the quick-start. The efficiency/deployment story leans hard on Alibaba's own infra (PAI-EAS, PAI-Blade, OSS file io) — outside that ecosystem you lose the inference optimization and deployment path the README is selling. Like mmdetection, you're learning EasyCV's own config DSL and hook system on top of PyTorch, which adds a debugging layer between you and the actual model code.