Hybrid-SORT

Hybrid-SORT is a simply and strong multi-object tracker.

Hybrid-SORT: Weak Cues Matter for Online Multi-Object Tracking

Abstract

Multi-Object Tracking (MOT) aims to detect and associate all desired objects across frames. Most methods accomplish the task by explicitly or implicitly leveraging strong cues (i.e., spatial and appearance information), which exhibit powerful instance-level discrimination. However, when object occlusion and clustering occur, both spatial and appearance information will become ambiguous simultaneously due to the high overlap between objects. In this paper, we demonstrate that this long-standing challenge in MOT can be efficiently and effectively resolved by incorporating weak cues to compensate for strong cues. Along with velocity direction, we introduce the confidence state and height state as potential weak cues. With superior performance, our method still maintains Simple, Online and Real-Time (SORT) characteristics. Furthermore, our method shows strong generalization for diverse trackers and scenarios in a plug-and-play and training-free manner. Significant and consistent improvements are observed when applying our method to 5 different representative trackers. Further, by leveraging both strong and weak cues, our method Hybrid-SORT achieves superior performance on diverse benchmarks, including MOT17, MOT20, and especially DanceTrack where interaction and occlusion are frequent and severe.

Highlights

Hybrid-SORT is a SOTA heuristic trackers on DanceTrack and performs excellently on MOT17/MOT20 datasets.
Maintains Simple, Online and Real-Time (SORT) characteristics.
Training-free and plug-and-play manner.
Strong generalization for diverse trackers and scenarios

Pipeline

News

[12/09/2023]: Hybrid-SORT is accepted by AAAI2024!
[08/24/2023]: Hybrid-SORT is supported in yolo_tracking. Many thanks to @mikel-brostrom for the contribution.
[08/01/2023]: The arxiv preprint of Hybrid-SORT is released.

Tracking performance

Results on DanceTrack test set

Tracker	HOTA	MOTA	IDF1	FPS
OC-SORT	54.6	89.6	54.6	30.3
Hybrid-SORT	62.2	91.6	63.0	27.8
Hybrid-SORT-ReID	65.7	91.8	67.4	15.5

Results on MOT20 challenge test set

Tracker	HOTA	MOTA	IDF1
OC-SORT	62.1	75.5	75.9
Hybrid-SORT	62.5	76.4	76.2
Hybrid-SORT-ReID	63.9	76.7	78.4

Results on MOT17 challenge test set

Tracker	HOTA	MOTA	IDF1
OC-SORT	63.2	78.0	77.5
Hybrid-SORT	63.6	79.3	78.4
Hybrid-SORT-ReID	64.0	79.9	78.7

Installation

Hybrid-SORT code is based on OC-SORT and FastReID. The ReID component is optional and based on FastReID. Tested the code with Python 3.8 + Pytorch 1.10.0 + torchvision 0.11.0.

Step1. Install Hybrid_SORT

git clone https://github.com/ymzis69/HybridSORT.git
cd HybridSORT
pip3 install -r requirements.txt
python3 setup.py develop

Step2. Install pycocotools.

pip3 install cython; pip3 install 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI'

Step3. Others

pip3 install cython_bbox pandas xmltodict

Step4. [optional] FastReID Installation

You can refer to FastReID Installation.

pip install -r fast_reid/docs/requirements.txt

Data preparation

Our data structure is the same as OC-SORT.

Download MOT17, MOT20, CrowdHuman, Cityperson, ETHZ, DanceTrack, CUHKSYSU and put them under <HYBRIDSORT_HOME>/datasets in the following structure (CrowdHuman, Cityperson and ETHZ are not needed if you download YOLOX weights from ByteTrack or OC-SORT) :

datasets
|——————mot
|        └——————train
|        └——————test
└——————crowdhuman
|        └——————Crowdhuman_train
|        └——————Crowdhuman_val
|        └——————annotation_train.odgt
|        └——————annotation_val.odgt
└——————MOT20
|        └——————train
|        └——————test
└——————Cityscapes
|        └——————images
|        └——————labels_with_ids
└——————ETHZ
|        └——————eth01
|        └——————...
|        └——————eth07
└——————CUHKSYSU
|        └——————images
|        └——————labels_with_ids
└——————dancetrack        
         └——————train
            └——————train_seqmap.txt
         └——————val
            └——————val_seqmap.txt
         └——————test
            └——————test_seqmap.txt

Prepare DanceTrack dataset:

# replace "dance" with ethz/mot17/mot20/crowdhuman/cityperson/cuhk for others
python3 tools/convert_dance_to_coco.py

Prepare MOT17/MOT20 dataset.

# build mixed training sets for MOT17 and MOT20 
python3 tools/mix_data_{ablation/mot17/mot20}.py

[optional] Prepare ReID datasets:

cd <HYBRIDSORT_HOME>

# For MOT17 
python3 fast_reid/datasets/generate_mot_patches.py --data_path <dataets_dir> --mot 17

# For MOT20
python3 fast_reid/datasets/generate_mot_patches.py --data_path <dataets_dir> --mot 20

# For DanceTrack
python3 fast_reid/datasets/generate_cuhksysu_dance_patches.py --data_path <dataets_dir>

Model Zoo

Download and store the trained models in 'pretrained' folder as follow:

<HYBRIDSORT_HOME>/pretrained

Detection Model

We provide some pretrained YOLO-X weights for Hybrid-SORT, which are inherited from ByteTrack.

Dataset	HOTA	IDF1	MOTA	Model
DanceTrack-val	59.3	60.6	89.5	Google Drive
DanceTrack-test	62.2	63.0	91.6	Google Drive
MOT17-half-val	67.1	78.0	75.8	Google Drive
MOT17-test	63.6	78.7	79.9	Google Drive
MOT20-test	62.5	78.4	76.7	Google Drive

For more YOLO-X weights, please refer to the model zoo of ByteTrack.

ReID Model

Ours ReID models for MOT17/MOT20 is the same as BoT-SORT , you can download from MOT17-SBS-S50, MOT20-SBS-S50, ReID models for DanceTrack is trained by ourself, you can download from DanceTrack.

Notes:

MOT20-SBS-S50 is trained by Deep-OC-SORT, because the weight from BOT-SORT is corrupted. Refer to Issue.
ReID models for DanceTrack is trained by ourself, with both DanceTrack and CUHKSYSU datasets.

Training

Train the Detection Model

You can use Hybrid-SORT without training by adopting existing detectors. But we borrow the training guidelines from ByteTrack in case you want work on your own detector.

Download the COCO-pretrained YOLOX weight here and put it under <HYBRIDSORT_HOME>/pretrained.

Train ablation model (MOT17 half train and CrowdHuman)

python3 tools/train.py -f exps/example/mot/yolox_x_ablation.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_x.pth

Train MOT17 test model (MOT17 train, CrowdHuman, Cityperson and ETHZ)

python3 tools/train.py -f exps/example/mot/yolox_x_mix_det.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_x.pth

Train MOT20 test model (MOT20 train, CrowdHuman)

For MOT20, you need to uncomment some code lines to add box clipping: [1],[2],[3] and [4]. Then run the command:
```
python3 tools/train.py -f exps/example/mot/yolox_x_mix_mot20_ch.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_x.pth
```

Train on DanceTrack train set

python3 tools/train.py -f exps/example/dancetrack/yolox_x.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_x.pth

Train custom dataset

First, you need to prepare your dataset in COCO format. You can refer to MOT-to-COCO or CrowdHuman-to-COCO. Then, you need to create a Exp file for your dataset. You can refer to the CrowdHuman training Exp file. Don't forget to modify get_data_loader() and get_eval_loader in your Exp file. Finally, you can train bytetrack on your dataset by running:
```
python3 tools/train.py -f exps/example/mot/your_exp_file.py -d 8 -b 48 --fp16 -o -c pretrained/yolox_x.pth
```

Train the ReID Model

After generating MOT ReID dataset as described in the 'Data Preparation' section.

cd <BoT-SORT_dir>

# For training MOT17 
python3 fast_reid/tools/train_net.py --config-file ./fast_reid/configs/MOT17/sbs_S50.yml MODEL.DEVICE "cuda:0"

# For training MOT20
python3 fast_reid/tools/train_net.py --config-file ./fast_reid/configs/MOT20/sbs_S50.yml MODEL.DEVICE "cuda:0"

# For training DanceTrack, we joint the CHUKSUSY to train ReID Model for DanceTrack
python3 fast_reid/tools/train_net.py --config-file ./fast_reid/configs/CUHKSYSU_DanceTrack/sbs_S50.yml MODEL.DEVICE "cuda:0"

Refer to FastReID repository for addition explanations and options.

Tracking

Notes:

Some parameters are set in the cfg.py. For example, if you run Hybrid-SORT on the dancetrack-val dataset, you should pay attention to the line 35-45 in exps/example/mot/yolox_dancetrack_val_hybrid_sort.py .
We set fp16==False on the MOT datasets becacuse fp16 will lead to significant result fluctuations.

DanceTrack

dancetrack-val dataset

# Hybrid-SORT
python tools/run_hybrid_sort_dance.py -f exps/example/mot/yolox_dancetrack_val_hybrid_sort.py -b 1 -d 1 --fp16 --fuse --expn $exp_name 

# Hybrid-SORT-ReID
python tools/run_hybrid_sort_dance.py -f exps/example/mot/yolox_dancetrack_val_hybrid_sort_reid.py -b 1 -d 1 --fp16 --fuse --expn $exp_name

dancetrack-test dataset

# Hybrid-SORT
python tools/run_hybrid_sort_dance.py --test -f exps/example/mot/yolox_dancetrack_test_hybrid_sort.py -b 1 -d 1 --fp16 --fuse --expn $exp_name

# Hybrid-SORT-ReID
python tools/run_hybrid_sort_dance.py --test -f exps/example/mot/yolox_dancetrack_test_hybrid_sort_reid.py -b 1 -d 1 --fp16 --fuse --expn $exp_name

MOT20

MOT20-test dataset

#Hybrid-SORT
python tools/run_hybrid_sort_dance.py -f exps/example/mot/yolox_x_mix_mot20_ch_hybrid_sort.py -b 1 -d 1 --fuse --mot20 --expn $exp_name 

#Hybrid-SORT-ReID
python tools/run_hybrid_sort_dance.py -f exps/example/mot/yolox_x_mix_mot20_ch_hybrid_sort_reid.py -b 1 -d 1 --fuse --mot20 --expn $exp_name

Hybrid-SORT is designed for online tracking, but offline interpolation has been demonstrated efficient for many cases and used by other online trackers. If you want to reproduct out result on MOT20-test dataset, please use the linear interpolation over existing tracking results:

# offline post-processing
python3 tools/interpolation.py $result_path $save_path

MOT17

MOT17-val dataset

# Hybrid-SORT
python3 tools/run_hybrid_sort_dance.py -f exps/example/mot/yolox_x_ablation_hybrid_sort.py -b 1 -d 1 --fuse --expn $exp_name 

# Hybrid-SORT-ReID
python3 tools/run_hybrid_sort_dance.py -f exps/example/mot/yolox_x_ablation_hybrid_sort_reid.py -b 1 -d 1 --fuse --expn  $exp_name

MOT17-test dataset

# Hybrid-SORT
python3 tools/run_hybrid_sort_dance.py -f exps/example/mot/yolox_x_mix_det_hybrid_sort.py -b 1 -d 1 --fuse --expn $exp_name

# Hybrid-SORT-ReID
python3 tools/run_hybrid_sort_dance.py -f exps/example/mot/yolox_x_mix_det_hybrid_sort_reid.py -b 1 -d 1 --fuse --expn $exp_name

Hybrid-SORT is designed for online tracking, but offline interpolation has been demonstrated efficient for many cases and used by other online trackers. If you want to reproduct out result on MOT17-test dataset, please use the linear interpolation over existing tracking results:

# offline post-processing
python3 tools/interpolation.py $result_path $save_path

Demo

Hybrid-SORT, with the parameter settings of the dancetrack-val dataset

python3 tools/demo_track.py --demo_type image -f exps/example/mot/yolox_dancetrack_val_hybrid_sort.py -c pretrained/ocsort_dance_model.pth.tar --path ./datasets/dancetrack/val/dancetrack0079/img1 --fp16 --fuse --save_result

Hybrid-SORT-ReID, with the parameter settings of the dancetrack-val dataset

python3 tools/demo_track.py --demo_type image -f exps/example/mot/yolox_dancetrack_val_hybrid_sort_reid.py -c pretrained/ocsort_dance_model.pth.tar --path ./datasets/dancetrack/val/dancetrack0079/img1 --fp16 --fuse --save_result

TCM on other trackers

download ReID weight from googlenet_part8_all_xavier_ckpt_56.h5 for MOTDT and DeepSORT.