Code release for "Detect to Track and Track to Detect", ICCV 2017
Clone or download
Latest commit e013785 Feb 20, 2018
Type Name Latest commit message Commit time
Failed to load latest commit information.
experiments init Jan 10, 2018
functions init Jan 10, 2018
imdb init Jan 10, 2018
models added network mean Jan 16, 2018
utils init Jan 10, 2018
.gitattributes init Jan 10, 2018
.gitignore added network mean Jan 16, 2018
.gitmodules init Jan 10, 2018
LICENSE init Jan 10, 2018 Update Feb 20, 2018
bbox_moments.mat added network mean Jan 16, 2018
download_proposals.m init Jan 10, 2018
get_root_path.m init Jan 10, 2018
rfcn_build.m init Jan 10, 2018
startup.m init Jan 10, 2018


Detect to Track and Track to Detect

This repository contains the code for our ICCV 2017 paper:

Christoph Feichtenhofer, Axel Pinz, Andrew Zisserman
"Detect to Track and Track to Detect"
in Proc. ICCV 2017
  • This repository also contains results for a ResNeXt-101 and Inception-v4 backbone network that perform slightly better (81.6% and 82.1% mAP on ImageNet VID val) than the ResNet-101 backbone (80.0% mAP) used in the conference version of the paper

  • This code builds on the original Matlab version of R-FCN

  • We are preparing a Python version of D&T that will support end-to-end training and inference of the RPN, Detector & Tracker.

If you find the code useful for your research, please cite our paper:

      title={Detect to Track and Track to Detect},
      author={Feichtenhofer, Christoph and Pinz, Axel and Zisserman, Andrew},
      booktitle={International Conference on Computer Vision (ICCV)},


The code was tested on Ubuntu 14.04, 16.04 and Windows 10 using NVIDIA Titan X or Z GPUs.

If you have questions regarding the implementation please contact:

Christoph Feichtenhofer <feichtenhofer AT>



  1. Download the code git clone --recursive
  • This will also download a modified version of the Caffe deep learning framework. In case of any issues, please follow the installation instructions in the corresponding README as well as on the Caffe website.
  1. Compile the code by running rfcn_build.m.

  2. Edit the file get_root_path.m to adjust the models and data paths.

    • Download the ImageNet VID dataset from
    • Download pretrained model files and the RPN proposals, linked below and unpack them into your models/data directory.
    • In case the models are not present, the function check_dl_model will attempt to download the model to the respective directories
    • In case the RPN files are not present, the function download_proposals will attempt to download & extract the proposal files to the respective directories


  • You can train your own models on ImageNet VID as follows
    • script_Detect_ILSVRC_vid_ResNet_OHEM_rpn(); to train the image-based Detection network.
    • script_DetectTrack_ILSVRC_vid_ResNet_OHEM_rpn(); to train the video-based Detection & Tacking network.


  • The scripts above have subroutines that test the learned models after training. You can also test our trained, final models available for download below. We provide three testing functions that work with a different numbers of frames at a time (i.e. processed by one GPU during the forward pass)
    1. rfcn_test(); to test the image-based Detection network.
    2. rfcn_test_vid(); to test the video-based Detection & Tacking network with 2 frames at a time.
    3. rfcn_test_vid_multiframe(); to test the video-based Detection & Tacking network with 3 frames at a time.
  • Moreover, we provide multiple testing network definitions that can be used for interesting experiments, for examüple
    • test_track.prototxt is the most simple form of D&T testing
    • test_track_reg.prototxt is a D&T version that additionally regresses the tracking boxes before performing the ROI tracking. Therefore, this procedure produces tracks that tightly encompass the underlying objects, whereas the above function tracks the proposal region (and therefore also the background area).
    • test_track_regcls.prototxt is a D&T version that additionally classifies the tracked region and computes the detection confidence as the mean of the detection score from the current frame, as well as the detection score of the tracked region in the next frame. Therefore, this method produces better results, especially if the temporal distance between the frames becomes larger and more complementary information can be integrated from the tracked region

Results on ImageNet VID

  • The networks are trained as decribed in the paper; i.e. on an intersection of the ImageNet object detection from video (VID) dataset which contains 30 classes in 3862 training videos and and the ImageNet object detection (DET) dataset (only using the data from the 30 VID classes). Validation results on the 555 videos of ImageNet VID validation are shown below.
Method test structure ResNet-50 ResNet-101 ResNeXt-101 Inception-v4
Detect test.prototxt 72.1 74.1 75.9 77.9
Detect & Track test_track.prototxt 76.5 79.8 81.4 82.0
Detect & Track test_track_regcls.prototxt 76.7 80.0 81.6 82.1
  • We show different testing network definitions in the rows and backbone networks in columns. The reported performance is mAP (in %), averaged over all videos and classes in the ImageNet VID validation subset.

Trained models


Our models were trained using region proposals extracted using a Region Proposal Network that is trained on the same data as D&T. We use the RPN from craftGBD and provide the extracted proposals for training and testing on ImageNet VID and the DET subsets below.

Pre-computed object proposals for