Skip to content

[ICRA2024]A fast and high-performance visual object tracker with real-time speed on Jetson.

License

Notifications You must be signed in to change notification settings

TsingWei/LiteTrack

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

21 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

LiteTrack

Our manuscript has been accepted by ICRA 2024.

Official implementation of [arxiv]"LiteTrack: Layer Pruning with Asynchronous Feature Extraction for Lightweight and Efficient Visual Tracking"

TODO list:

  • Release training and testing code.
  • Release our trained models.
  • Release onnx conversion and inference code.

Performance

Accuracy FPS
GOT-10k LaSOT TrackingNet 2080Ti OrinNX XavierNX OrinNX(onnx) XavierNX(onnx)
LiteTrack-B9-3 72.2 67 82.4 171 21.4 16.2 64.75 37.15
LiteTrack-B8-3 70.4 66.4 81.4 190 24.81 18.1 69.97 40.07
LiteTrack-B6-3 68.7 64.6 80.8 237 31.5 20.5 82.53 50.61
LiteTrack-B4-2 65.2 62.5 79.9 315 44.1 25.3 101.98 66.94
HiT-B 64.0 64.6 80.0 175(Not Tested)
OSTrack 71.0 69 83.1 140 17.99 12.08 50.72 30.89

Prepare Environment

Ensure that you have install the pytorch >= 1.8.1 and corresponding torchvision version. It doesn't matter whether you use pip or conda.

Then execute

bash install.sh

You can check the script where I install the opencv-python-headless library because of my headless server environment. If you are running in a machine with GUI, you can comment that out.

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 ./output

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

Training and Evaluation on Datasets

We support training following OSTrack.

Dataset Preparation

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

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

Training

Download pre-trained CAE ViT-Base weights(cae_base.pth) and put it under $PROJECT_ROOT$/pretrained_models.

NOTE: ViT in CAE is slightly different from the original ones (in Image is worth 16x16 words and MAE), e.g., projections of Q,K,V and layer scale. Details can be seen in the code.

Run the command below to train the model:

# single GPU
python tracking/train.py --script litetrack --config B8_cae_center_got10k_ep100 --save_dir ./output --mode single  --use_wandb 0

# 2 GPUs
python tracking/train.py --script litetrack --config B8_cae_center_got10k_ep100 --save_dir ./output --mode multiple --nproc_per_node 2  --use_wandb 0
  • If you want to use wandb to record detailed log and get clear curve of training, set --use_wandb 1.
  • The batch size, learning rate and the other detailed parameters can be set in config file, e.g., experiments/litetrack/B6_cae_center_got10k_ep100.yaml.
  • In our implemention, we always follow batch_size $\times$ num_GPU $\div$ learning_rate $= 32 \times 1 \div 0.0001$ for alignment with OSTrack. For example, if u use 4 GPUs and bs=32 (in each GPU), then the lr should be 0.0004.
  • If you are using an Ampere GPU (RTX 30X0), we suggest you upgrade pytorch to 2.x to get free boost on attention computation, which will save about half of the GPU memory so that u can enable batch size up to 128 in one single RTX 3090 (it costs <19 G memory with AMP).
  • We save the checkpoints without optimization params in the last 20% epochs, for testing over these epochs to avoid accuracy jittering.

Evaluation

Use your own training weights or ours(BaiduNetdisk:lite or Google Drive) in $PROJECT_ROOT$/output/checkpoints/train/litetrack.
Some testing examples:

  • LaSOT (cost ~2 hours on 2080Ti with i5-11400F for one epoch testing)
python tracking/test.py litetrack B6_cae_center_all_ep300 --dataset lasot --threads 8 --num_gpus 1 --ep 300 299 290
python tracking/analysis_results.py # need to modify tracker configs and names

For other off-line evaluated benchmarks, modify --dataset correspondingly.

  • GOT10K-test (cost ~7 mins on 2080Ti with i5-11400F for one epoch testing)
python tracking/test.py litetrack B6_cae_center_got10k_ep100 --dataset got10k_test --threads 8 --num_gpus 1 --ep 100 99 98
python lib/test/utils/transform_got10k.py --tracker_name litetrack --cfg_name B6_cae_center_got10k_ep100_099 # the last number is epoch
  • TrackingNet (cost ~1 hour on 2080Ti with i5-11400F for one epoch testing)
python tracking/test.py litetrack B6_cae_center_all_ep300 --dataset got10k_test --threads 8 --num_gpus 1 --ep 300 299
python lib/test/utils/transform_trackingnet.py --tracker_name litetrack --cfg_name B6_cae_center_all_ep300_300 # the last number is epoch

Acknowledgement

Our code is built upon OSTrack. Also grateful for PyTracking.

About

[ICRA2024]A fast and high-performance visual object tracker with real-time speed on Jetson.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published