search

MVIG-SJTU / AlphaPose

Real-Time and Accurate Full-Body Multi-Person Pose Estimation&Tracking System
http://mvig.org/research/alphapose.html
Other
7.94k stars 1.97k forks source link

What does "gt mAP", "rcnn mAP", " gt box" and "det box" means? #967

Closed thisimyusername closed 2 years ago

thisimyusername commented 2 years ago

I have tried train some models(hrnet,fastpose and simeplepose) using custom datasets, and obtain many models.When training it print: Epoch 153 | gt mAP: 0.6215162432424949 | rcnn mAP: 0.6274187838001606 and when run validate.py with "model_153.pth", It print: gt box: 0.6333331405098032 mAP | det box: 0.595698565478996 mAP. I have much confusion with "gt mAP", "rcnn mAP", " gt box" and "det box".

dekathomas commented 2 years ago

Hello, sorry maybe I can't answer your question, but I have the problem when I try train using FastPose model and I use Google Colab when doing training I have you have time, please check my issue. I need your help I you want to help me.

Thanks in advance!

HaoyiZhu commented 2 years ago

Alphapose uses top-down methods which will first detect human bboxes and then detect human keypoints. So the gt mAP and the gt box means using groundtruth bboxes, while rcnn mAP and det box means using bboxes generated by detectors (default is yolo v3).

dekathomas commented 2 years ago

Hello @thisimyusername, where is the place you can change the model used by training algorithm?

HaoyiZhu commented 2 years ago

@dekathomas Hi, if you want to use a custom model, you can check alphapose/models/. We use FastPose by default.

dekathomas commented 2 years ago

owh sorry @HaoyiZhu I mean the base network (hrnet,fastpose and simeplepose), because by default alphapose use resnet right? i want to try mobilenet for speed

HaoyiZhu commented 2 years ago

@dekathomas Ah, I think you can find it under alphapose/models/layers/. The backbones such as resnet are there.

dekathomas commented 2 years ago

sorry for late reply @HaoyiZhu, I wanna try mobilenet, but I don't find mobilenet in layers folder. And if I look at the pytorch documentation, torchvision 0.3.0 doesn't support mobilenet. I have tried to upgrade the torchvision version, but I got an error when trying to run the training. Do you have any suggestions for me to be able to use mobilenet?

HaoyiZhu commented 2 years ago

I see. Actually we do not support mobilenet yet. Maybe you can implement it by yourself, or you can update yout pytorch version. I think higher version of pytorch can support higher version of torchvision？ @dekathomas

dekathomas commented 2 years ago

Hmm oke I will try it later. Thanks for your information.

dekathomas commented 2 years ago

by the way @HaoyiZhu why you give the name rcnn mAP if the validation is by yolo bounding box?

HaoyiZhu commented 2 years ago

I think it is a history remain? @Fang-Haoshu