Skip to content

Latest commit

 

History

History
88 lines (60 loc) · 2.96 KB

README.md

File metadata and controls

88 lines (60 loc) · 2.96 KB
Raw

Datasets for StableKeypoints

User guide of the 5 datasets used in StableKeypoints:

  1. CelebA
  2. CUB
  3. DeepFashion
  4. Taichi
  5. Human3.6m

Background

Recently, I came across an intriguing project called StableKeypoints, which was published at CVPR 2024.

I attempted to run the project using its dataset but encountered some issues with almost every dataset.

After searching for solutions on the website and experimenting with the code, I managed to resolve most of these problems.

To make it more convenient for myself and others in the future, I am compiling some useful resources here.

Time Consuming for optimization embedding of 500 steps

Different datasets almost have the same time consuming for optimization embedding. Because the iteration numbers is num_steps * (batch_size // num_gpus), which means the dataset size doesn't affect the iteration numbers.

The default batch_size is 4, so num_gpus must less than or equal to 4. Commonly, num_gpus is 4, so the total iterations are num_steps * 4 // 4 = num_steps. num_steps default is 500.

In practical experiments, the time consuming for optimization embedding about 20~30 minutes, affected by the status of the gpu (busy or idle).

  • celeba: 19 minutes
  • cub: 31 minutes
  • deepfashion: 19 minutes
  • human36m: 20 minutes
  • taichi: 32 minutes

dataset size

Dataset Train Test
taichi 5000 300
cub_aligned 5964 2874
cub_001 29 17
cub_002 30 14
cub_003 30 15
cub_all 5964 2874
deepfashion 10604 1179
celeba_aligned 19000 1000
celeba_wild 5379 283
human3.6m 796648 87975
unaligned_human3.6m 159444 17615

Time Consuming for Precomputing keypoints

The iteration depends on min(len(dataset), max_num_points).

The dataset here means the training set. And max_num_points is 50000.

Different datasets have different sizes, so the iteration numbers may vary.

Dataset Time Images
taichi 2 hours 5000
cub 4 hours 5964
deepfashion 7 hours 10604
celeba 9.3 hours 19000
human36m 22 hours 50000 (actually 796648)

Time Consuming for evaluating

The iteration is exactly on len(test dataset).

So it will totally depends on the size of the test dataset.

Dataset Time Images
taichi 0.1 hours 300
celeba 0.5 hours 1000
deepfashion 1 hour 1179
cub 2 hours 2874
human36m - 87975