Adapting Object Detectors

Implementation of the paper Adapting Object Detectors from Images to Weakly Labeled Videos.

Usage

The first step of our approach is to generate a shortlist of object proposals from source and target images. We use the edge boxes algorithm for generating the object proposals. Let K be the number of object proposals generated on the image. We represent each proposal as a 4096-dimensional CNN feature vector.

Run the experiment

1. Get into each folder named as class name [i.e. 01_aeroplane].
2. Keep the .t7 file of the dataset with data, ground truth bounding box, class label and proposals.
3. Run doall.lua file using th command [i.e. th ../doall.lua].
4. After training, run evaluate.sh file for getting the corloc for each class.

Model Architecture

Citation

If you find this project useful for your work, please consider cite the paper.

@article{BMVC2017Adapt,
  author = {Omit Chanda, Eu Wern Teh, Mrigank Rochan, Zhenyu Guo and Yang Wang},
  title = {Adapting Object Detectors from Images to Weakly Labeled Videos},
  journal = {The 28th British Machine Vision Conference (BMVC), 2017},
  year = {2017}
}

Others

1. This code is initially based on @soumith's code but most of the things are modified.
2. For more information please refer to the paper.