A Self supervised Learning System for Object Detection in Videos Using Random Walks on Graphs

Overall

This code is a pytorch implementation of the unsupervised object detection method introduced in our paper "A Self supervised Learning System for Object Detection in Videos Using Random Walks on Graphs".

Paper:

https://arxiv.org/abs/2011.05459\

Video

Requirements

• Python 3.7
• pytorch 1.5
• torchvision 0.6
• cuda 10

Training and evaluation

We provide an example on CORe50 datasets: \

1. Download the datasets "cropped_128x128_images.zip" on https://vlomonaco.github.io/core50/index.html#download
2. Under the project root RWS/, run:

   source setup.sh
   

3. Under RWS/script/CORe50/data_processing/, provide data path in the data_processing.sh file. Then, run:

   sh data_processing.sh
   

4. Under RWS/script/CORe50/graph_construction/ run:

   sh build_graph.sh
   

   To visualize the similarity graph, run:

   python graph_viz
   

5. Under RWS/script/CORe50/train/, run:

   sh random_walk_sampling.sh
   

   Then, run:

   sh train.sh
   

6. To evaluate the trained projection layers, under RWS/script/CORe50/evaluate/ run:

   sh unsupervised_cluster.sh
   

Reference

Is you find the method useful, please consider cite the paper:

@misc{tan2020selfsupervised,
      title={A Self-supervised Learning System for Object Detection in Videos Using Random Walks on Graphs}, 
      author={Juntao Tan and Changkyu Song and Abdeslam Boularias},
      year={2020},
      eprint={2011.05459},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}