MCITrack

• [AAAI'2025] - Exploring Enhanced Contextual Information for Video-Level Object Tracking

Exploring Enhanced Contextual Information for Video-Level Object Tracking
Ben Kang, Xin Chen, Simiao Lai, Yang Liu, Yi Liu, Dong Wang

[Models] , [raw_results], [training_logs]

This is an official pytorch implementation of the paper Exploring Enhanced Contextual Information for Video-Level Object Tracking.

Highlights

New contextual information propagation method

MCITrack utilizes hidden states to efficiently transmit richer and more significant contextual information.

Simple architecture

MCITrack has a simple structure, consisting of a backbone, a contextual information fusion module, and a prediction head.

Strong performance

Comparsion with SOTA Models:

Tracker	LaSOT (AUC)	LaSOT_ext (AUC)	TrackingNet (AUC)	GOT10K (A0)
MCITrack-B224	75.3	54.6	86.3	77.9
ODTrack-B384	73.2	52.4	85.1	77.0
ARTrackV2-256	71.6	50.8	84.9	75.9
LoRAT-B224	71.7	50.3	83.5	72.1

Large-Scale Comparsion:

Tracker	LaSOT (AUC)	LaSOT_ext (AUC)	TrackingNet (AUC)	GOT10K (A0)
MCITrack-L384	76.6	55.7	87.9	80.0
ODTrack-L384	74.0	53.9	86.1	78.2
ARTrackV2-L384	73.6	53.4	86.1	79.5
LoRAT-L378	75.1	56.6	85.6	77.5

Install the environment

conda create -n mcitrack python=3.11
conda activate mcitrack
bash install.sh

• Add the project path to environment variables

export PYTHONPATH=<absolute_path_of_MCITrack>:$PYTHONPATH

Data Preparation

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

${MCITrack_ROOT}
 -- data
     -- lasot
         |-- airplane
         |-- basketball
         |-- bear
         ...
     -- got10k
         |-- test
         |-- train
         |-- val
     -- coco
         |-- annotations
         |-- images
     -- trackingnet
         |-- TRAIN_0
         |-- TRAIN_1
         ...
         |-- TRAIN_11
         |-- TEST
     -- vasttrack
         |-- Zither
         |-- Zebra
         ...

Set project paths

Run the following command to set paths for this project

python tracking/create_default_local_file.py --workspace_dir . --data_dir ./data --save_dir .

After running this command, you can also modify paths by editing these two files

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

Train

Download pre-trained weights and put it under ./pretrained

Train MCITrack

torchrun --nproc_per_node 8 lib/train/run_training.py --script mcitrack --config mcitrack_b224 --save_dir .

Test and evaluate on benchmarks

Put the downloaded checkpoints under ./checkpoints/train/mcitrack

• LaSOT

python tracking/test.py mcitrack mcitrack_b224 --dataset lasot --threads 2
python tracking/analysis_results.py # need to modify tracker configs and names

• LaSOT_ext

python tracking/test.py mcitrack mcitrack_b224 --dataset lasot_extension_subset --threads 2
python tracking/analysis_results.py # need to modify tracker configs and names

• GOT10K-test

python tracking/test.py mcitrack mcitrack_b224_got --dataset got10k_test --threads 2
python lib/test/utils/transform_got10k.py --tracker_name mcitrack --cfg_name mcitrack_b224_got

• TrackingNet

python tracking/test.py mcitrack mcitrack_b224 --dataset trackingnet --threads 2
python lib/test/utils/transform_trackingnet.py --tracker_name mcitrack --cfg_name mcitrack_b224

• TNL2K

python tracking/test.py mcitrack mcitrack_b224 --dataset tnl2k --threads 2
python tracking/analysis_results.py # need to modify tracker configs and names

• UAV123

python tracking/test.py mcitrack mcitrack_b224 --dataset uav --threads 2
python tracking/analysis_results.py # need to modify tracker configs and names

• NFS

python tracking/test.py mcitrack mcitrack_b224 --dataset nfs --threads 2
python tracking/analysis_results.py # need to modify tracker configs and names

Test FLOPs, Params and Speed

python tracking/profile_model.py --script mcitrack --config mcitrack_b224

Citation

@inproceedings{kang2025exploring,
  title={Exploring Enhanced Contextual Information for Video-Level Object Tracking}, 
  author={Ben Kang and Xin Chen and Simiao Lai and Yang Liu and Yi Liu and Dong Wang},
  booktitle={AAAI},
  year={2025}
}

Contact

• Ben Kang (email:[email protected])