Skip to content
Branch: master
Find file History
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
data update init Jun 3, 2019
experiments update init Jun 3, 2019
external update init Jun 3, 2019
fetch_data update init Jun 3, 2019
functions update init Jun 3, 2019
imdb update init Jun 3, 2019
models update init Jun 3, 2019
utils update init Jun 3, 2019 update init Jun 3, 2019 Update Jun 6, 2019
mspld_build.m update init Jun 3, 2019
startup.m update init Jun 3, 2019

Few-Example Object Detection with Model Communication

By Xuanyi Dong, Liang Zheng, Fan Ma, Yi Yang, Deyu Meng


This project aims to solve the few-example object detection problem, in which there are only a few images with bounding box annotations per object class and a large number of images without annotations. We propose an algorithm to combine the self-paced learning and the multi-modal learning, and we call it Multi-modal Self-Paced Learning for Detection (MSPLD). The implementation is modified from R-FCN and Faster RCNN.

Note that few-example object detection is a special case of semi-supervised object detection. However, most works on semi-supervised object detection assume that some classes have many strong bounding boxes, while others have weak image-level labels. A brief comparison between MSPLD and weakly supervised/semi-supervised/few-example object detection is shown below:

Please refer to the paper for more detailed comparison.


If you find MSPLD useful in your research, please consider citing:

  title   = {Few-Example Object Detection with Model Communication},
  author  = {Dong, Xuanyi and Zheng, Liang and Ma, Fan and Yang, Yi and Meng, Deyu},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
  volume  = {41},
  number  = {7},
  pages   = {1641-1654},
  doi     = {10.1109/TPAMI.2018.2844853}, 
  ISSN    = {0162-8828}, 
  year    = {2018}


Requirements: software

  • Caffe build for MSPLD (included in this repository, see external/caffe)
  • MATLAB 2014a or later

Data Preparation

  • Download PASCAL VOC 2007 and PASCAL VOC 2012. Follow the README in datasets
  • Download the region proposals extracted by Selective Search or EdgeBox.
  • Download the pre-trained Imagenet models for ResNet, VGG, and GoogleNet

Training & Testing

  1. Run experiments/VOC07_Tr_Res50E_Res101S_VGG16F to repoduce the results on VOC 2007.


  1. Selective Search Data: Google Drive
  2. YFCC100M Data for Ablative Study: Google Drive
  3. Detection models are in the models directory.
You can’t perform that action at this time.