PromptVT

This project is the implementation of paper PromptVT, including models, raw results, and testing codes(the training code will be uploaded after organizing).

❗ Ubuntu(Linux) is highly recommended, Windows has some weird installation problems and model inference problems.

❗ This is the CPU edition, no CUDA or GPU required.

Performance

PromptVT achieves SOTA performance on 8 benchmarks in lightweight trackers.

Usage

Installation

Create and activate a conda environment:

conda create -n PromptVT python=3.7
conda activate PromptVT

Install the required packages:

bash install_PromptVT.sh

Data Preparation

${PromptVT_ROOT}
 -- data
     -- lasot
         |-- airplane
         |-- basketball
         |-- bear
         ...
     -- got10k
         |-- test
         |-- train
         |-- val
     -- OTB100
         |-- Basketball
         |-- Biker
         ...
     -- trackingnet
         |-- TRAIN_0
         |-- TRAIN_1
         ...
         |-- TRAIN_11
         |-- TEST
     -- uav123
         |-- anno
              |-- UAV123
         |-- data_seq
              |-- UAV123
     -- Anti-UAV
         |-- Test
              |-- 20190925_111757_1_1
              ...
     -- Anti-UAV-410
         |-- Test
              |-- 02_6319_1500-2999
              ...
         

Path Setting

Run the following command to set paths:

cd < PATH_of_PromptVT >
python tracking/create_default_local_file.py --workspace_dir . --data_dir ./data --save_dir .

You can also modify paths by these two files:

./lib/train/admin/local.py  # paths for training
./lib/test/evaluation/local.py  # paths for testing

Test and evaluate PromptVT on benchmarks

If you want to use ONNX model, set ' use_onnx = True ' in ./lib/test/tracker/PromptVT.py.

• LaSOT

  python tracking/test.py --tracker_name PromptVT --tracker_param baseline --dataset lasot
  python tracking/analysis_results.py # need to modify tracker configs and names
  

• GOT10K-test

  python tracking/test.py --tracker_name PromptVT --tracker_param baseline --dataset got10k_test
  python lib/test/utils/transform_got10k.py --tracker_name PromptVT --cfg_name baseline
  

  Upload the results to the official GOT-10K evaluation server.

• TrackingNet

  python tracking/test.py --tracker_name PromptVT --tracker_param baseline --dataset trackingnet
  python lib/test/utils/transform_trackingnet.py --tracker_name PromptVT --cfg_name baseline
  

  Upload the results to the official TrackingNet evaluation server.

• UAV123

  python tracking/test.py --tracker_name PromptVT --tracker_param baseline --dataset uav
  python tracking/analysis_results.py # need to modify tracker configs and names
  

• AntiUAV

  python tracking/test.py --tracker_name PromptVT --tracker_param baseline --dataset antiuav
  python tracking/analysis_results.py # need to modify tracker configs and names
  

  The raw data is labeled in json format, which we converted to OTB-like-txt format to fit our tracking library. The converted file is located at . /tracking/AntiUAVJSON2OTBTxt.py.

• AntiUAV410

  python tracking/test.py --tracker_name PromptVT --tracker_param baseline --dataset antiuav410
  python tracking/analysis_results.py # need to modify tracker configs and names
  

  The raw data is labeled in json format, which we converted to OTB-like-txt format to fit our tracking library. The converted file is located at . /tracking/AntiUAVJSON2OTBTxt.py.

• OTB100

  python tracking/test.py --tracker_name PromptVT --tracker_param baseline --dataset otb
  python tracking/analysis_results.py # need to modify tracker configs and names
  

• VOT2020
  modify the path sets in ./external/vot20/trackers.ini, ./lib/test/vot20/PromptVT.py, and ./lib/test/vot20/PromptVT_vot20.py.

  cd external/vot20/PromptVT
  bash exp.sh
  

Test FLOPs, Params, and FPS

FLOPs and Params:

modify the ' yaml_fname ' in ./tracking/profile_model.py.

python tracking/profile_model.py

FPS:

place the tracking/Calculate_FPS.py in the tracking results folder and run it.

Model Zoo & Raw Results

The trained models and the raw tracking results are provided in the model zoo.
put PyTorch model and ONNX model in ./checkpoints/PromptVT/baseline/.
We also provide model conversion scripts./tracking/****_onnx.py.

Acknowledgments

Thanks for the PyTracking and STARK for helping us quickly implement our ideas.

Contact

If you have any question, feel free to email [email protected]. ^_^