RepNet PyTorch
GitHub repository: https://github.com/materight/RepNet-pytorch.
A PyTorch port with pre-trained weights of RepNet, from Counting Out Time: Class Agnostic Video Repetition Counting in the Wild (CVPR 2020) [paper] [project] [notebook].
This repo provides an implementation of RepNet written in PyTorch and a script to convert the pre-trained TensorFlow weights provided by the authors. The outputs of the two implementations are almost identical, with a small deviation (less than $10^{-6}$ at most) probably caused by the limited precision of floating point operations.
Get Started
- Clone this repo and install dependencies:
git clone https://github.com/materight/RepNet-pytorch
cd RepNet-pytorch
pip install -r requirements.txt
- To download the TensorFlow pre-trained weights and convert them to PyTorch, run:
python convert_weights.py
Run inference
Simply run:
python run.py
The script will download a sample video, run inference on it and save the count visualization. You can also specify a video path as argument (either a local path or a YouTube/HTTP URL):
python run.py --video_path [video_path]
If the model does not produce good results, try to run the script with more stride values using --strides
.
Example of generated videos showing the repetition count, with the periodicity score and the temporal self-similarity matrix: