chen-si-jia / track_AMMC

AMMC(Augmentation by Mimicking Motion Change)

This is the official implementation of our ACM MM 2021 oral paper Do We Really Need Frame-by-Frame Annotation Datasets for Object Tracking?

Setup AMMC

• Install AMMC

git clone https://github.com/wsumel/AMMC-Augmentation-by-Mimicking-Motion-Change-.git
cd  pytracking

      Run the install script to install the dependencies.You need to provide the ${conda_install_path} (e.g. ~/anaconda3) and the name ${env_name} for the created conda environment (e.g. pytracking).

# install dependencies
bash install.sh ${conda_install_path} ${env_name}
conda activate pytracking

• Download AMMC Models

     Pretrain model ATOM + AMMC and Dimp + AMMC can be download from Baidu key:ammc

• Run tracker

conda activate pytracking
cd ./pytracking
# ATOM
python run_tracker.py atom default --dataset_name dataset_name --sequence sequence --debug debug --threads threads
# Dimp
python run_tracker.py dimp dimp50 --dataset_name dataset_name --sequence sequence --debug debug --threads threads

How to reproduce our result

• Download FAT(Few-annotation Tracking) benchmark

     Baidu key:ammc

• Training the tracker

conda activate pytracking
# ATOM + AMMC
python run_training.py bbreg atom_default
# Dimp + AMMC
python run_training.py dimp dimp50

How to construct your own FAT benchmark

• Download three base datasets, namely TrackingNet, GOT-10k, LaSOT

• Setting the inital size of the FAT: fat_frame_number on ./ltr/admin/local.py

• Run the script file and Modify the dataset path as your own

cd ./ltr/
python extract_dataset.py

Performance

• LaSOT

  Tracker	Success Score	Precision Score
  ATOM (paper)	0.515	n/a
  ATOM (ammc 1)	0.476	0.465
  ATOM (ammc 3)	0.507	0.501
  ATOM (ammc 5)	0.510	0.505
  ATOM (ammc 10)	0.514	0.510
  ATOM (ammc 1-10 best)	0.514	0.510
  ATOM (ammc paper)	0.517	n/a
  DiMP50 (paper)	0.569	n/a
  DiMP50 (ammc 1)	0.524	0.502
  DiMP50 (ammc 3)	0.562	0.553
  DiMP50 (ammc 5)	0.571	0.567
  DiMP50 (ammc 10)	0.568	0.558
  DiMP50 (ammc 1-10 best)	0.571	0.567
  DiMP50 (ammc paper)	0.569	n/a

• GOT-10k:

  Tracker	Success Score (AO)	SR(0.50)	SR(0.75)
  ATOM (paper)	0.556	0.634	0.402
  ATOM (ammc 1)	0.515	0.614	0.311
  ATOM (ammc 3)	0.553	0.645	0.408
  ATOM (ammc 5))	0.551	0.641	0.409
  ATOM (ammc 10)	0.549	0.636	0.408
  ATOM (ammc 1-10 best)	0.553	0.645	0.408
  ATOM (ammc paper)	0.564	0.661	0.411
  DiMP50 (paper)	0.611	0.717	0.492
  DiMP50 (ammc 1)	0.525	0.626	0.323
  DiMP50 (ammc 3)	0.581	0.680	0.446
  DiMP50 (ammc 5)	0.605	0.708	0.488
  DiMP50 (ammc 10)	0.615	0.722	0.486
  DiMP50 (ammc 1-10 best)	0.615	0.722	0.486
  DiMP50 (ammc paper)	0.622	0.731	0.494

• TrackingNet:

  Tracker	Success Score	Precision Score	Normalized Precision Score
  ATOM (paper)	0.703	0.648	0.771
  ATOM (ammc 1)	0.675	0.618	0.751
  ATOM (ammc 3)	0.712	0.650	0.770
  ATOM (ammc 5)	0.716	0.654	0.775
  ATOM (ammc 10)	0.713	0.650	0.768
  ATOM (ammc 1-10 best)	0.716	0.654	0.775
  ATOM (ammc paper)	0.712	0.648	0.769
  DiMP50 (paper)	0.740	0.687	0.801
  DiMP50 (ammc 1)	0.698	0.635	0.765
  DiMP50 (ammc 3)	0.740	0.682	0.795
  DiMP50 (ammc 5)	0.742	0.681	0.796
  DiMP50 (ammc 10)	0.747	0.685	0.797
  DiMP50 (ammc 1-10 best)	0.747	0.685	0.797
  DiMP50 (ammc paper)	0.746	0.687	0.797

• OTB-100/OTB-2015:

  Tracker	Success Score	Precision Score
  ATOM (paper)	0.669	n/a
  ATOM (ammc 1)	0.662	0.885
  ATOM (ammc 3)	0.649	0.865
  ATOM (ammc 5)	0.638	0.858
  ATOM (ammc 10)	0.643	0.850
  ATOM (ammc 1-10 best)	0.662	0.885
  ATOM (ammc paper)	0.669	n/a
  DiMP50 (paper)	0.684	n/a
  DiMP50 (ammc 1)	0.652	0.862
  DiMP50 (ammc 3)	0.647	0.852
  DiMP50 (ammc 5)	0.643	0.850
  DiMP50 (ammc 10)	0.640	0.847
  DiMP50 (ammc 1-10 best)	0.652	0.862
  DiMP50 (ammc paper)	0.669	n/a

• UAV

  Tracker	Success Score	Precision Score
  ATOM (paper)	0.644	n/a
  ATOM (ammc 1)	0.622	0.850
  ATOM (ammc 3)	0.624	0.832
  ATOM (ammc 5)	0.607	0.816
  ATOM (ammc 10)	0.620	0.837
  ATOM (ammc 1-10 best)	0.624	0.832
  ATOM (ammc paper)	0.640	n/a
  DiMP50 (paper)	0.654	n/a
  DiMP50 (ammc 1)	0.628	0.844
  DiMP50 (ammc 3)	0.643	0.856
  DiMP50 (ammc 5)	0.619	0.820
  DiMP50 (ammc 10)	0.622	0.832
  DiMP50 (ammc 1-10 best)	0.643	0.856
  DiMP50 (ammc paper)	0.667	n/a

Acknowledgments

This repo is based on Pytracking which is an excellent work