Realtime Object Recognition on the COCO dataset using YOLOv4.
Example made with love by Jonathan Frank 2022
Video credits to Videvo
Image credits to pytorch-YOLOv4 (or wherever they got it...)
Model provided by https://github.com/AlexeyAB/darknet
Model converted by https://github.com/hunglc007/tensorflow-yolov4-tflite
The model was trained on the COCO dataset. The dataset consists of 80 everyday classes. Here is an excerpt of things it is able to recognize:
- person
- bicycle
- car
- bird
- cat
- umbrella
- handbag
- frisbee
- bottle
- wine glass
- fork
- spoon
- orange
- pizza
- ...
For the full list please check out bin/data/cocoClasses.txt after downloading the example model.
This example comes with a converted model. If you are interested in the approach of conversion feel free to check out this repository.
The model expects an image as input and outputs candidate regions for the objects of interests. Using a method called Non-maximum Suppresion (NMS) we can filter the proposals. For more information on NMS please check this blogpost.