microsoft / voxelpose-pytorch

Official implementation of "VoxelPose: Towards Multi-Camera 3D Human Pose Estimation in Wild Environment"
MIT License
486 stars 92 forks source link

Deciding on SPACE_SIZE and SPACE_CENTER #10

Closed egeozsoy closed 3 years ago

egeozsoy commented 3 years ago

I have a custom dataset that I want to apply voxelpose to. But there is no explanation for how the values SPACE_SIZE and SPACE_CENTER were selected for the 3 datasets, therefore it is not clear to me what to set them to for my custom dataset

CHUNYUWANG commented 3 years ago

It depends on the calibration results. The basic idea is to calculate a 3D bounding box which covers the whole motion space. Then SPACE_CENTER denotes the center coordinate of the box and SPACE_SIZE denotes the lengths of the box.

egeozsoy commented 3 years ago

I see thank you. In that case making this box a little bit bigger than it needs to be should not degrade the performance right?

CHUNYUWANG commented 3 years ago

That's right. But you may need to consider the size of each voxel (which depends on the box size and number of voxels) as that affects the quantization error.

StephanPan commented 3 years ago

hi, @egeozsoy .What's your trainsets and testsets? Your custom datasets or open-source datasets? Did you implement the cross-datasets evaluation.

sudo-vinnie commented 3 years ago

@egeozsoy
@CHUNYUWANG

I am also facing the same issue with a custom datasets? Could you guys make a simple explanation the details regarding the parameters for:

 MULTI_PERSON:
  SPACE_SIZE:
    - 12000.0
    - 12000.0
    - 20000
  SPACE_CENTER:
    - 3000.0
    - 4500.0
    - 1000.0
  INITIAL_CUBE_SIZE:
    - 80
    - 80
    - 20
#  MAX_PEOPLE_NUM: 1
  THRESHOLD: 0.3
PICT_STRUCT:
  GRID_SIZE:
    - 2000.0
    - 2000.0
    - 2000.0
  CUBE_SIZE:
    - 64
    - 64
    - 64