search

chensong1995 / HybridPose

HybridPose: 6D Object Pose Estimation under Hybrid Representation (CVPR 2020)
MIT License
419 stars 64 forks source link

train with my own data #93

Open kordou opened 10 months ago

kordou commented 10 months ago

HI,

i am new on machine learning, i am strated to study Ai this uear and i have a eersize to train a model to recognise to 6dof pose of an object, (eg a cup of coffe)

the idea is to use synthetic data from a 3d engine to create the dataset

can i somehow create the dataset fro this model ?

thank you in advanced for your help

chensong1995 commented 10 months ago

Hi kordou,

Thanks for your interest in our work!

You can create a synthetic dataset to train HybridPose. My recommendation is to start by examining the provided datasets (i.e., Linemod, Blender Linemod, Fuse Linemod). You may want to pay attention to how keypoints, symmetry correspondences, and segmentation masks are labeled. The "Dataset set-up" section in README provides some good guidelines.

However, creating a custom dataset for HybridPose (our work) can be a bit tricky for someone who is just starting to study AI. This is because HybridPose requires a lot of intermediate representation labels. Perhaps, starting with a slightly simpler approach can smooth the learning curve a bit. I think PVNet (our baseline approach) is a good alternative because it only requires one type of intermediate representation (the keypoints).

I hope this helps! Let me know if you have further concerns.

kordou commented 10 months ago

Hi chensong1995, thank you for your answer. it is very helpfull