Given an arbitrary multi-person image，how to obtain the pose keypoint estimation?

NieXC / pytorch-ppn

Pytorch implementation of Pose Partition Networks for Multi-Person Pose Estimation (ECCV'18)

MIT License

90 stars 20 forks source link

Closed janicevidal closed 1 year ago

janicevidal commented 5 years ago

RT, without bbox and person center information?

NieXC commented 5 years ago

For arbitrary input image, you can just set its objpos, scale_provided and center_box to be None, then call the same evaluation function to produce multi-person pose estimation results. Details can be referred to the following code provided in the evaluation function: https://github.com/NieXC/pytorch-ppn/blob/1b79e0f7b62243f2474e0bc4af8930fe97332aa1/utils/eval_util.py#L118-L125

The hyper-parameter dist_th for hierarchical clustering may need to be adjusted depending on the size of your input image.

janicevidal commented 5 years ago

@NieXC how could the training process working for other dataset，like COCO？Have you ever tried?

NieXC commented 5 years ago

@janicevidal

For other datasets, you just need to modify the data_loader according to their joint annotations, then follow the same training procedure.

(Noted: the Local Greedy Search in inference phase needs to be adjusted accordingly if the joint hierarchy is modified.)

I have tried to train and evaluate our model on MSCOCO dataset, the code is under cleanning, which will be updated in this repo lately.