Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time

Target Driven Instance Detection

This is an implementation of the technique described in Target Driven Instance Detection. It is written in python for use with Pytorch.

External Requirements


These instructions will setup the code and data to run our experiments on the AVD dataset. More instructions will be provided to run our other experiments or use your own data.

  1. Dependencies and Data:
  • Make sure you have Pytorch (and torchvision)
  • Get the AVD processing code, and make sure it is included in your PYTHONPATH
  • Download the AVD Data into a path of your choosing, we will refer to is as AVD_ROOT_DIR.
  • Make sure to also get the instance id map and put it in the AVD_ROOT_DIR
  • Download the target images into a path of your choosing, we will refer to is as TARGET_IMAGE_DIR.
  1. Get the code
git clone
  1. Install the other requirements
cd target_driven_instance_detection/
pip install -r requirements.txt
  1. Build the cython code for anchor boxes and non-max supression
cd model_defs/
  1. Build the coco evaluation cython code
cd ../evaluation/cocoapi/PythonAPI/
make all
cd ../../../
  1. Convert AVD annotations to COCO format yourself, or download the converted files

To Download the files:

mkdir Data
cd Data

Download the tar here

tar -xf tdid_gt_boxes.tar

Or to convert yourself:

cd  evaluation/
#Update paths in `` with:
#a path to save the annotations, we will call it VAL_GROUND_TRUTH_BOXES

#now update the scene_list in 
#to make the test set
#change the path to save the annotations, we will call it TEST_GROUND_TRUTH_BOXES
  1. Set paths configs/ file. See configs/ for details on config files. Make sure to update the config with your:

  2. Start training!

#make sure you are in root directory of project, target_driven_instance_detection/

Trained models

Here are models trained for each of the 3 splits on the AVD dataset


Please cite our paper if you find our work useful:

  title = {Target Driven Instance Detection},
  author = {Ammirato, Phil, and Fu, Cheng-Yang and Shvets, Mykhailo and Kosecka, Jana and Berg, Alexander C.},
  booktitle = {arXiv:1803.04610},
  year = {2018}


Things to clean and add

  1. Add data and configs for GMU to AVD experiment
  2. Add data and configs for RGB-D Scenes one-shot classifcation experiment
  3. Check det4class code
  4. Clean eval by object
  5. Provide trained models
  6. make a note about downloading pretrained pytorch models
  7. How to add your own data

Improvements to system

  1. How to choose target image, multiview targt image pooling thing


This code started as a modification of a Faster-RCNN Pytorch implementation here, and still uses some of that code. (In particular the nms code).


No description, website, or topics provided.




No releases published


No packages published