torch-hog 0.0.1

Non Official PyTorch Implementation of the HoG Descriptor

torch-hog

Non Official PyTorch Implementation of the HoG Descriptor

Recent updates:

• 0.0.1: Basic HoG calcuation was added

Contents

• Installation
• Basic Use
• Citing
• Repository and discussions

Installation

Plain and simple:

pip install torch-hog

Basice Use

Import

Use as a PyTorch Module:

from torch_hog import HoG

hog = HoG(num_bins=9, cell_size=8, padding="reflect")

Or as a function:

from torch_hog.functional import hog

Dense Feature Extraction

either way, the API is the same. To get the features at every possible coordinate, just feed in the batch:

# Create a random batch of images
B, C, H, W = 5, 3, 360, 480
img = torch.randn((B, C, H, W))

dense_hog_features = hog(img) # returns a tensor of shape (B, H*W, bins)

Sparse Feature Extraction

You can specify the desired coordinates for every image in the batch.

# specify the desired coordinates
coords = torch.tensor([
  [20, 15],
  [17.7, 18.9], # yes we're accepting floats too
  [150, 170],
])

# Create a random batch of images
B, C, H, W = 5, 3, 360, 480
img = torch.randn((B, C, H, W))

hog_features = hog(img, coords=coords) # returns a tensor of shape (B, 3, bins)

You can also specify different number of coordinates for every image:

# specify the desired coordinates
coords1 = torch.tensor([
  [20, 15],
  [17.7, 18.9], # yes we're accepting floats too
  [150, 170],
])

coords2 = torch.tensor([
  [41, 38],
  [69, 85],
])
coords = [coords1, coords2]

# Create a random batch of images
B, C, H, W = 5, 3, 360, 480
img = torch.randn((B, C, H, W))

hog_features = hog(img, coords=coords) # returns a list of tensors as with the shapes: [(3, bins), (2, bins)]

Cite Us

Please use the following bibtex record if you're using this project in your research:

@misc{TorchHog2025,
  author = {Abe, Amit},
  title = {Torch-HoG},
  year = {2025},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/wamitw/torch-hog}},
}