search

NVlabs / Deep_Object_Pose

Deep Object Pose Estimation (DOPE) – ROS inference (CoRL 2018)
Other
1.03k stars 287 forks source link

need help --custom object training #232

Closed HiHiAllen closed 2 years ago

HiHiAllen commented 2 years ago

HI, i have some doubt about my custom object training .Using NDDS, i generated 20k images to train the net and run it 38 epochs (from epoch 20-38 , the loss fluctuates around 0.0018 ， so i shut it down ) Then i input several images from those 20k images , and i got their beliefs maps ,it seems quite good ,i strongly believe that the training works. Here are some belief maps: 选区_064 选区_055 选区_056 选区_059 选区_058

BUT ,when i input some images photographed by my camera (astra pro 640*480), target cant be identified . Here are some belief maps:

选区_065 选区_066 选区_067

so,whats your opinion , do i need more images to train the net ? is 50k enough? or some else problems? Thanks for any help.

TontonTremblay commented 2 years ago

It looks like there might be a symmetry on the object. I cannot see very well from my phone. But looking at the output from the belief maps I would think there is a sym.

mintar commented 2 years ago

On the PC, it doesn't look like there is a rotational symmetry, only a mirror symmetry which should be fine.

HiHiAllen commented 2 years ago

thanks , i wil try to train a new one to make a comparison.