// the find
WZMIAOMIAO/deep-learning-for-image-processing
deep learning for image processing including classification and object-detection etc.
A teaching repo that walks through implementing classic CV architectures from scratch in both PyTorch and TensorFlow2 — LeNet through ConvNeXt and MobileViT for classification, Faster-RCNN/YOLO/FCOS for detection, FCN/DeepLabV3/U-Net for segmentation, plus HRNet and DeepPose for keypoints. Best fit is someone who wants to actually understand how ResNet or Vision Transformer works by reading clean, from-scratch code rather than importing a model from a library. Paired bilibili video walkthroughs (in Mandarin) explain the papers and the code side by side.
27k stars and 8k forks, actively pushed into 2026, so it's still maintained rather than a dead student project.
Each architecture gets its own self-contained folder with model.py, train.py, predict.py, and usually a README — you can clone just pytorch_classification/ConvNeXt and run it without pulling in the whole repo's dependency graph. Covers the full CV pipeline, not just classification: detection, segmentation, instance segmentation, and keypoint estimation all have working from-scratch implementations, which is unusual breadth for one repo. The deploying_service folder is a genuinely useful extra — ONNX export, OpenVINO, TensorRT, and a Flask serving example, so it doesn't stop at training like most tutorial repos.
Almost entirely gated behind bilibili video explanations — the written docs are mostly link lists to videos, and the videos are Mandarin-only with no English subtitles mentioned, which locks out a large chunk of the audience the stars suggest it has. No tests, no CI, and no requirements.txt at the repo root — dependency versions are scattered per-subfolder and inconsistent, so expect pinning fights getting an older folder like Test1_official_demo running next to a newer one like MobileViT. These are training/reference implementations, not production code: no distributed training abstractions beyond a few manual multi_GPU scripts, no config system, and model weights/checkpoints are mostly left as an exercise (external download links) rather than versioned releases.