search

chenyilun95 / tf-cpn

Cascaded Pyramid Network for Multi-Person Pose Estimation (CVPR 2018)
MIT License
792 stars 196 forks source link

About data preprocess and data augmentation #19

Closed xiashh closed 6 years ago

xiashh commented 6 years ago

I want to know more detail about data prepocessing especially cropping strategy.

chenyilun95 commented 6 years ago

You can find the summary of our data preprocessing and post-processing strategies in our paper.

xiashh commented 6 years ago

Do you use bounding box of ground truth directly without no disturbance during training?

chenyilun95 commented 6 years ago

Multi-scale training is applied

xiashh commented 6 years ago

Yes, I see that in dataset.py. I wonder know why you do not use results of the detector and calculate IOU as the prediction of instance.

chenyilun95 commented 6 years ago

I didn’t get it. If you mean we directly train using the detection boxes, I think its performance works like multi scale training and maybe random shifts of the boxes. I think random shifts of boxes can help, which we didn’t try it. As well if you want to exploit the specific detector.