UVLTrack

The official implementation of our AAAI 2024 paper Unifying Visual and Vision-Language Tracking via Contrastive Learning

Brief Introduction

Single object tracking aims to locate the target object in a video sequence according to the state specified by different modal references, including the initial bounding box (BBOX), natural language (NL), or both (NL+BBOX). Different previous modality-specific trackers, we present a unified tracker called UVLTrack, which can simultaneously handle all three reference settings (BBOX, NL, NL+BBOX) with the same parameters, allowing wider appliation scenarios. The proposed UVLTrack enjoys several merits. First, we design a modality-unified feature extractor for joint visual and language feature learning and propose a multi-modal contrastive loss to align the visual and language features into a unified semantic space. Second, a modality-adaptive box head is proposed, which makes full use of the target reference to mine ever-changing scenario features dynamically from video contexts and distinguish the target in a contrastive way, enabling robust performance in different reference settings.

Strong Performance

UVLTrack presents strong performance under different reference settings.

Reference Modality	NL	BBOX	NL+BBOX
Methods	TNL2K (AUC / P)	TNL2K(AUC / P)	TNL2K(AUC / P)
OSTrack-256	-/-	54.3/-	-/-
OSTrack-384	-/-	55.9/-	-/-
VLTTT	-/-	-/-	53.1/53.3
JointNLT	54.6/55.0	-/-	56.9/58.1
UVLTrack-B	55.7/57.2	62.7/65.4	63.1/66.7
UVLTrack-L	58.2/60.9	64.8/68.8	64.9/69.3

Raw results can be downloaded from Google Drive or Baidu Netdisk.

Checkpoints

You can download the model weights from Google Drive or Baidu Netdisk.

Methods	UVLTrack-B	UVLTrack-L
Reference Modality	TNL2K (AUC / P)	TNL2K(AUC / P)
NL	55.0/56.8	58.2/60.9
BBOX	63.0/66.3	65.2/69.5
NL+BBOX	62.9/66.5	64.8/69.2

Install the environment

Use the Anaconda

conda env create -f uvltrack_env.yaml

or

conda create -n uvltrack python=3.6
conda activate uvltrack
bash install.sh

Data Preparation

Put the tracking datasets in ./data. It should look like:

${UVLTrack_ROOT}
-- data
-- lasot
|-- airplane
|-- basketball
|-- bear
...
-- lasotext
|-- atv
|-- badminton
|-- cosplay
...
-- got10k
|-- test
|-- train
|-- val
-- coco
|-- annotations
|-- train2017
-- trackingnet
|-- TRAIN_0
|-- TRAIN_1
...
|-- TRAIN_11
|-- TEST
-- otb99
|-- OTB_query_test
|-- OTB_query_train
|-- OTB_videos
-- refcocog
|-- refcocog
|-- split # download this folder from VLTVG (https://github.com/yangli18/VLTVG/blob/master/docs/get_started.md)
|-- train2014 # coco 2014
|-- val2014 # coco 2014
-- tnl2k
|-- test
|-- train

Train UVLTrack

Download the pretrained MAE and BERT, put it under <PROJECT_ROOT>/pretrain.

Training with multiple GPUs using DDP.

# UVLTrack
sh scripts/train.sh uvltrack baseline_base

Evaluation

Download the model weight from Google Drive or Baidu Netdisk.

Put the downloaded weights on <PROJECT_ROOT>/checkpoints/train/uvltrack/baseline_base and <PROJECT_ROOT>/checkpoints/train/uvltrack/baseline_large correspondingly.

Notably, the modality of target reference (NL, BBOX or NLBBOX) is specified in config TEST.MODE

# Testing
sh scripts/test.sh uvltrack baseline_base <dataset_name> <num_threads_per_gpu> <num_gpu>
# Evaluation
python tracking/analysis_results.py --tracker_name uvltrack --tracker_param baseline_base --dataset_name <dataset_name>_<reference_modality>_<EPOCH>
# Example
sh scripts/test.sh uvltrack baseline_base otb99 4 2
python tracking/analysis_results.py --tracker_name uvltrack --tracker_param baseline_base --dataset_name otb99_NL_300

Run UVLTrack on your own video

Specify the target by bounding box or natural language, which should keep consistent with TEST.MODE in config.

sh scripts/demo.sh uvltrack baseline_base \
<input video path> \
<output video path> \
<language description of target> \
<initial bbox of target: x y w h>

Test Speed

# UVLTrack-B: About 60 FPS on NVIDIA RTX 3090 GPU
python tracking/profile_model.py --script uvltrack --config baseline_base
# UVLTrack-L: About 34 FPS on NVIDIA RTX 3090 GPU
python tracking/profile_model.py --script uvltrack --config baseline_large

Contact

For questions about our paper or code, please contact Yinchao Ma or Yuyang Tang

Acknowledgments

• Thanks for JointNLT, Mixformer and OSTrack Library, which helps us to quickly implement our ideas.

• We use the implementation of the ViT from the Timm repo and BERT from the pytorch_pretrained_bert.

Citation

If our work is useful for your research, please consider cite:

@misc{ma2024unifying,
title={Unifying Visual and Vision-Language Tracking via Contrastive Learning}, author={Yinchao Ma and Yuyang Tang and Wenfei Yang and Tianzhu Zhang and Jinpeng Zhang and Mengxue Kang},
year={2024},
eprint={2401.11228},
archivePrefix={arXiv},
primaryClass={cs.CV}
}