yunyang1994/tensorflow-yolov3

This repository is a Python implementation of YOLOv3, an algorithm that detects and labels objects in photos and videos. Given an image, the model can identify dozens of different types of things, such as people, cars, bicycles, and animals, and draw a box around each one. The underlying paper is titled "YOLOv3: An Incremental Improvement" and this code translates that research into a working program built with TensorFlow, a popular machine-learning framework. To get started, you clone the repository, install the listed dependencies, then download a set of pre-trained weights. Those weights represent everything the model already learned from a large public dataset called COCO, which contains images tagged across 80 common categories. After converting the weights to a format TensorFlow can use, you can run either image_demo.py or video_demo.py to see object detection in action immediately, without any additional training. The project also supports training on your own image collection. You prepare a text file listing your images along with bounding-box coordinates and category labels for each object, then optionally start from the pre-trained COCO weights (the recommended path) rather than training from scratch. An evaluate script and a separate mAP calculator let you measure how accurately the model performs on a test set. The author notes at the top of the README that a newer version built on TensorFlow 2.0 exists in a separate repository, along with a Chinese-language blog post explaining how the detector works from scratch. The code here targets the older TensorFlow 1.x API. This is research code aimed at people who want to experiment with object detection or adapt the YOLOv3 architecture to their own datasets. It is not a packaged library with a stable API, and the README is primarily a setup guide rather than full documentation.