Why is it about mmpose, because I am a dish. This article is only for record use, that is to say, what is written may not be correct. Welcome to criticize and correct.
I just want to know how this kind of key point detection is marked and how the data set is made. I can’t find it on the whole network. I don’t deserve the paid resources. Today is 12.30, and it will be New Year’s Day. It will be sent out before New Year’s Day!
mmpose Chinese Tutorial: Tutorial 1: How to fine-tune the model — MMPose 0.29.0 documentation
1. Label
If it is to label a single target, use labelme, and use coco-annotate for multiple targets. For my task, this article only introduces the use of labelme for labeling.
Open the labelme, first draw a large frame, select the Create Rectangle option under Edit, frame the location that needs to be marked, and then select the Create Point option, click key points in this box, each point cannot be the same, save and generate a json file , can not be used directly, the generated json file needs to be converted into a coco json file.
Open the tools/json2json.py file for conversion and separate training and validation.
2. run
There are a lot of environment configurations on the Internet, just do it. When training your own data set, there is a data set registration mechanism. First imitate the original version and write the animal you want to detect. You need to modify the name and the number of key points, and then in the animal Import the newly created file in the /_init_.py file.
Insert a new class in apis/inference.py, line 252, 253, line 688-693, line 689 is to find out the points that need to be connected, and manually connect them together.
Modify the pre-training weight, number of key points, batch size, and various paths in configs/animal/resnet/fish/res152_fish_192x192.py. After the configuration is complete, you can directly train in tools/train.py.
It is almost the same process, and then it can be divided into attitude detection with detection frame and posture detection without detection frame. This is the posture detection without frame, just imitate the fly, if you want to have a detection frame, imitate the horse.