DepthTrack: Cluster Meets BEV for Multi-Camera Multi-Target 3D Tracking

Team SKKU-AutoLab - ID 15

Track 1: Multi-Camera 3D Perception of Automation Lab at Sungkyunkwan University

This paper has been accepted and will be presented at the ICCV 2025 Workshop.

---

I. Dataset preparation

a. Data download

Go to the website of AI-City Challenge to get the dataset.

• https://www.aicitychallenge.org/2025-track1/

Download dataset to the folder <MTMC_Tracking_2025>

The dataset folder structure should be as following:

<MTMC_Tracking_2025>
│   ├── test
│   │   ├── Warehouse_017
│   │   │   ├── videos
│   │   │   ├── depth_maps
│   │   │   ├── calibration.json
│   │   │   └── map.png
│   │   ├── Warehouse_018
│   │   ├── Warehouse_019
│   │   └── Warehouse_020
...

b. Data enhance download:

Download the data enhance files from the link below and put them in the folder <MTMC_Tracking_2025>/test/Warehouse_018/videos to replace the original.

• https://drive.google.com/drive/folders/1ny4Co3uV9gL-kJngGJn-4HoGztgtFGvT?usp=drive_link

<MTMC_Tracking_2025>
│   ├── test
│   │   ├── Warehouse_017
│   │   ├── Warehouse_018
│   │   │   ├── videos
│   │   │   │   ├── Camera_0000.mp4
│   │   │   │   ├── ...
│   │   │   │   └── Camera_0008.mp4

---

II. Environment setup

a. Installation Miniconda or Anaconda:

1. Download & install Miniconda or Anaconda from https://docs.conda.io/projects/conda/en/latest/user-guide/install/linux.html

b. Create conda environment:

Follow the instructions in the following files to install the required dependencies.

• setup/setup_aic25_track1.sh
• setup/setup_kpr.sh
• setup/setup_mmpose.sh
• setup/setup_tracking.sh

AIC25_Track_01
│   ├── setup
│   │   ├── setup_aic25_track1.sh
│   │   ├── setup_kpr.sh
│   │   ├── setup_mmpose.sh
│   │   └── setup_tracking.sh

c. Load weights:

Downnoad each weight file and move them into the corresponding folder:

• https://drive.google.com/drive/folders/1AsPDJ1CWYkqO5Njg1lBC8HwIWTI-1Hq5?usp=sharing

Move folder aic25_track1 into:

• model_zoo/aic25_track1

AIC25_Track_01
│   ├── model_zoo
│   │   └── aic25_track1

Move weight swin_base_patch4_window7_224_22k.pth into:

• ultilities/pose_reid_preprocess/keypoint_promptable_reidentification/pretrained_models/SOLIDER/swin_base_patch4_window7_224_22k.pth

AIC25_Track_01
│   ├── ultilities
│   │   ├── pose_reid_preprocess
│   │   │   ├── keypoint_promptable_reidentification
│   │   │   │   ├── pretrained_models
│   │   │   │   │   ├── SOLIDER
│   │   │   │   │   │   ├── swin_base_patch4_window7_224_22k.pth

Move weight td-hm_hrnet-w48_8xb32-210e_coco-256x192-0e67c616_20220913.pth into:

• ultilities/pose_reid_preprocess/td-hm_hrnet-w48_8xb32-210e_coco-256x192-0e67c616_20220913.pth

AIC25_Track_01
│   ├── ultilities
│   │   ├── pose_reid_preprocess
│   │   │   ├── td-hm_hrnet-w48_8xb32-210e_coco-256x192-0e67c616_20220913.pth

---

III. Inference

a. Adjust the configuration

Default link to the dataset in the code is /media/vsw/Data1/MTMC_Tracking_2025/. You need to change it to your own path.

Change the varible ROOT_DATA_FOLDER in the file ultilities/pose_reid_preprocess/configuration.py to the folder path of the dataset <MTMC_Tracking_2025>.

• ROOT_DATA_FOLDER = <MTMC_Tracking_2025> (example ROOT_DATA_FOLDER ='/media/vsw/Data1/MTMC_Tracking_2025/')

Change the varible ROOT_DATA_FOLDER in the file ultilities/tracking/configuration.py to the folder path of the dataset <MTMC_Tracking_2025>.

• ROOT_DATA_FOLDER = <MTMC_Tracking_2025> (example ROOT_DATA_FOLDER ='/media/vsw/Data1/MTMC_Tracking_2025/')

Change the varible FOLDER_ROOT in the file ultilities/configuration.py to the folder path of the folder contain dataset <MTMC_Tracking_2025>.

• FOLDER_ROOT = dirname(<MTMC_Tracking_2025>) (example FOLDER_ROOT = '/media/vsw/Data1/')
• FOLDER_DATA_VERSION = "MTMC_Tracking_2025"

b. Run the code

Run the following commands in the terminal:

# Run all commands below from the folder AIC25_TRack_01

conda activate aic25_track1

bash run_inference_series.sh

cd ultilities/pose_reid_preprocess
conda activate mmpose
python run_pose.py 
cd ../..

cd ultilities/pose_reid_preprocess/keypoint_promptable_reidentification
conda activate kpr
python run_kpr_aicity25.py
cd ../../..

cd ultilities/tracking
conda activate tracking
python sv_tracking.py
python mv_tracking_prepare.py
python mv_tracking.py
cd ../..

conda activate aic25_track1
python ultilities/mapping_3d.py
python ultilities/filter_objects_out_bev.py
    

After running all command above, the output files will be in the folder

• <MTMC_Tracking_2025>/ExtractFrames/lookup_table/final_result_filtered.txt

<MTMC_Tracking_2025>
│   ├── ExtractFrames
│   │   ├── lookup_table
│   │   │   └── final_result_filtered.txt

---

IV. Training (optional)

Training dataset:

• https://drive.google.com/drive/folders/1ZETSfAbYNwVBoc_GBjWz3mjYhp2HgNRt?usp=sharing

---

V. Acknowledgement

Most of the code is adapted from Mon.

This repository also features code from Ultralytics, Torchreid, MMPose and Bot-Sort