Multi-Label Image Classification via Knowledge Distillation from Weakly-Supervised Detection

This repository contains the code (in Caffe) for the paper:

Multi-Label Image Classification via Knowledge Distillation from Weakly-Supervised Detection [ACM DL] [arXiv]
Yongcheng Liu, Lu Sheng, Jing Shao*, Junjie Yan, Shiming Xiang and Chunhong Pan
ACM Multimedia 2018

Project Page: https://yochengliu.github.io/MLIC-KD-WSD/

Weakly Supervised Detection (WSD)

• We use WSDDN¹ as the detection model, i.e., the teacher model.

• Because the released code of WSDDN is implemented using Matlab (based on MatConvNet), we first reproduce this paper using Caffe.

[1]. Hakan Bilen, Andrea Vedaldi, "Weakly Supervised Deep Detection Networks". In: IEEE Computer Vision and Pattern Recognition, 2016.

Reproduction results

detection

• Paper

    training: 5 scales + mirror          testing: fusion of 5 scales + mirror
  

• Our

    training: 5 scales + mirror          testing: single-forward test
  

classification

Datalist Preparation

image_path one_hot_label_vector(e.g., 0 1 1 ...) proposal_info(e.g., x_min y_min x_max y_max score x_min y_min x_max y_max score ...)

Training & Test

    ./wsddn/wsddn_train(deploy).prototxt

• VGG16 is used as the backbone model.

• For training, we did not use spatial regularizer. More details can be referred in the paper.

• For testing, you can use Pycaffe or Matcaffe.

Multi-Label Image Classification (MLIC)

• The MLIC model in our framework, i.e., the student model, is very compact for efficiency.

• It is constituted by a popular CNN model (VGG16, as the backbone model) following a fully connected layer (as the classifier).

• The backbone model of the student could be different from the teacher's.

Cross-Task Knowledge Distillation

Stage 1: Feature-Level Knowledge Transfer

    ./kd/train_stage1.prototxt

Stage 2: Prediction-Level Knowledge Transfer

    ./kd/train_stage2.prototxt

Datalist preparation is the same as mentioned in WSD. More details can be referred in our paper.

Implementation

Please refer to caffe-MLIC for details.

Citation

If our paper is helpful for your research, please consider citing:

    @inproceedings{liu2018mlickdwsd,   
      author = {Yongcheng Liu and    
                Lu Sheng and    
                Jing Shao and   
                Junjie Yan and   
                Shiming Xiang and   
                Chunhong Pan},   
      title = {Multi-Label Image Classification via Knowledge Distillation from Weakly-Supervised Detection},   
      booktitle = {ACM International Conference on Multimedia},    
      pages = {700--708},  
      year = {2018}   
    }   

Contact

If you have some ideas or questions about our research to share with us, please contact [email protected]