Skip to content

Latest commit

 

History

History
122 lines (110 loc) · 7.14 KB

GET_STARTED.md

File metadata and controls

122 lines (110 loc) · 7.14 KB
Raw

Get Started

Our data structure is the same as OC-SORT.

Data preparation

1. Datasets

Download MOT17, MOT20, CrowdHuman, Cityperson, ETHZ, DanceTrack and put them under <OCSORT_HOME>/datasets in the following structure:

datasets
|——————MOT17
|        └——————train
|        └——————test
└——————crowdhuman
|        └——————Crowdhuman_train
|        └——————Crowdhuman_val
|        └——————annotation_train.odgt
|        └——————annotation_val.odgt
└——————MOT20
|        └——————train
|        └——————test
└——————Cityscapes
|        └——————images
|        └——————labels_with_ids
└——————ETHZ
|        └——————eth01
|        └——————...
|        └——————eth07
└——————dancetrack        
        └——————train
        └——————val
        └——————test

2. YOLOX-Detection-Output

Download MOT17, MOT20, DanceTrack preprocessing detection frame file under detector YOLOX and put them under <BAM_HOME>/exps in the following structure:

exps
|——————dancetrack
|        └——————yolox_x
|           └——————val
|           └——————test
|——————MOT17
|        └——————yolox_x
|           └——————test
|        └——————ablation
|           └——————val
|——————MOT20
|        └——————yolox_x
|           └——————test

3. YOLOX and ReID Output

Please download the per-frame detection results and corresponding ReID feature vectors for each bounding box of the MOT17、MOT20、DanceTrack, processed with the YOLOX detector. Place the files in the <BAM_HOME>/exps directory, maintaining the following structure:

exps
|——————reid
|        └——————dancetrack
|           └——————yolox_x
|               └——————test
|        └——————MOT17
|           └——————yolox_x
|               └——————test
|        └——————MOT20
|           └——————yolox_x
|               └——————test

Evaluation

online

  • on DanceTrack Val set

    python tools/run_bamsort.py --det_type yolox_x --dataset dancetrack --dataset_type val --w_bec 0.3 --bec_num 4 --min_hits 7 --std_time_since_update 5 --std_switch_cnt 1 --std_max_hits 50 --fp16 --fuse --expn $exp_name

    We follow the TrackEval protocol for evaluation on the officially released validation set. This gives HOTA = 56.5 ~ 56.9.

  • on DanceTrack Test set

    python tools/run_bamsort.py --det_type yolox_x --dataset dancetrack --dataset_type test --w_bec 0.3 --bec_num 4 --min_hits 5 --std_time_since_update 5 --std_switch_cnt 1 --std_max_hits 50 --expn $exp_name

  • on DanceTrack Test set with ReID

    python tools/run_bamsort_reid.py --sort_with_reid --det_type yolox_x --dataset dancetrack --dataset_type test --w_bec 0.3 --bec_num 4 --min_hits 5 --std_time_since_update 5 --std_switch_cnt 1 --std_max_hits 50 --expn $exp_name

  • on MOT17 half val

    python tools/run_bamsort.py --det_type ablation --dataset MOT17 --dataset_type val --w_bec 0.4 --bec_num 4 --min_hits 1 --sort_with_std False --expn $exp_name

    We follow the TrackEval protocol for evaluation on the self-splitted validation set. This gives you HOTA = 67.3 ~ 67.7.

  • on MOT17/MOT20 Test set

    # MOT17
python tools/run_bamsort.py --det_type yolox_x --dataset MOT17 --dataset_type test --w_bec 0.4 --bec_num 4 --min_hits 1 --sort_with_std False --expn $exp_name 

# MOT20
python tools/run_bamsort.py --det_type yolox_x --dataset MOT20 --dataset_type test --w_bec 0.3 --bec_num 6 --min_hits 1 --sort_with_std False --expn $exp_name

    Following the adaptive detection thresholds by ByteTrack can further boost the performance.

  • on MOT17/MOT20 Test set with ReID

    # MOT17
python tools/run_bamsort_reid.py --sort_with_reid --det_type yolox_x --dataset MOT17 --dataset_type test --w_bec 0.4 --bec_num 4 --min_hits 1 --sort_with_std False --expn $exp_name 

# MOT20
python tools/run_bamsort_reid.py --sort_with_reid --det_type yolox_x --dataset MOT20 --dataset_type test --w_bec 0.3 --bec_num 6 --min_hits 1 --sort_with_std False --expn $exp_name

    Following the adaptive detection thresholds by ByteTrack can further boost the performance.

Note: We find the current implementation may show some randomness in different running trials. We are still inspecting this.

offline Interpolation

I have utilized the interpolation features from the OC-SORT algorithm to post-process existing tracking results. The OC-SORT algorithm offers two interpolation methods: linear interpolation and Gaussian Process Regression interpolation. To use linear interpolation on the results output by BAM-SORT online tracking, you can run the following command:

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

The OC-SORT algorithm also provides an attempt to use Gaussian Process Regression interpolation, which operates on existing linear interpolation results. To use Gaussian Process Regression interpolation, you can run the following command:

    python tools/gp_interpolation.py $raw_results_path $linear_interp_path $save_path

For the benchmark results of DanceTrack, please submit your output files to the DanceTrack evaluation site. Following submission, you can expect a HOTA score ranging from 64.0 to 64.5. For the benchmark tests of MOT17 and MOT20, outputs should be submitted to MOTChallenge. In this case, the expected HOTA score for MOT17 is approximately 64.5, while for MOT20 it is around 62.0.