Detect-Track is the official implementation of the ICCV 2017 paper Detect to Track and Track to Detect by Christoph Feichtenhofer, Axel Pinz, and Andrew Zisserman. The framework unifies object detection and tracking into a single pipeline, allowing detection to support tracking and tracking to enhance detection performance. Built upon a modified version of R-FCN, the code provides implementations using backbone networks such as ResNet-50, ResNet-101, ResNeXt-101, and Inception-v4, with results demonstrating state-of-the-art accuracy on the ImageNet VID dataset. The repository includes MATLAB-based training and testing scripts, along with pre-trained models and pre-computed region proposals for reproducibility. Multiple testing configurations are available, including multi-frame input and enhanced versions that refine tracking boxes and integrate detection confidence across frames.
Features
- Implements Detect-to-Track and Track-to-Detect framework (ICCV 2017)
- Built on a modified R-FCN with ResNet, ResNeXt, and Inception backbones
- Provides pre-trained models and pre-computed region proposals
- Training and testing scripts for ImageNet VID and DET datasets
- Multiple testing modes including multi-frame and refined tracking
- Results achieve over 82% mAP on ImageNet VID validation set
Project Samples
