search

asanakoy / deeppose_tf

DeepPose implementation on TensorFlow. Original Paper http://arxiv.org/abs/1312.4659
Other
144 stars 61 forks source link

How do I use my own images for prediction? #4

Closed shiangyoung closed 7 years ago

shiangyoung commented 7 years ago

Hello,

I was wondering if I could use my own images for prediction. I tried to run the test_snapshot.py. I'm looking for a result similar to #2 where the images are visualized. Could someone help me out?

Thank you.

asanakoy commented 7 years ago

@shiangyoung just get the predictions out of the network and write a script for visualization. You will ned to save predicted_joints . Yo can check evaluate_pcp() to see how predicted joints are calculated.

shiangyoung commented 7 years ago

Hi @asanakoy,

We are currently running some tests however the predicted joints that we receive back is inaccurate. Could I ask if we are required to annotate the images before using them for testing?

Thank you for your earlier feedback.

asanakoy commented 7 years ago

@shiangyoung as I understood you right, for testing you need annotations only to get a PCP/PCKh score. You need to train until the net converges.

shiangyoung commented 7 years ago

Hi @asanakoy,

Yup, we are currently training and it has reached around 220k steps. From what I understand, it would then be possible to use our own image for prediction without any prior annotations? We are actually aiming to get a visualization of the prediction using our own image.

Thanks.