Skip to content

Latest commit

 

History

History
84 lines (77 loc) · 4.38 KB

README.md

File metadata and controls

84 lines (77 loc) · 4.38 KB
Raw

MCMOT: One-shot multi-class multi-object tracking

单阶段实时多类别多目标跟踪
This is an extention work of FairMOT, which extends the one-class multi-object tracking to multi-class multi-object tracking
You can refer to origin fork FairMOT

Tracking demo of C5(car, bicycle, person, cyclist, tricycle)

image image image image image

Tracking demo of VISDrone dataset

image image image image image image

VisDrone dataset training with 10 object classes

VisDrone link
VisDrone is a public dataset for 4 CV challenges: object detection, crowd counting, single class multi-object tracking, multi-class multi-object tracking.

  • Download multi-class multi-object tracking part of Visdrone dataset
  • Using gen_dataset_visdrone.py script to generate labels.
  • Call the gen_dot_train_file function in gen_dataset_visdrone.py to generate the dot train file for VisDrone mcmot training task.
  • Uncomment cls2id and id2cls in multitracker.py to use the correct class names and class Ids mapping.
from gen_dataset_visdrone import cls2id, id2cls  # visdrone
# from gen_labels_detrac_mcmot import cls2id, id2cls  # mcmot_c5
  • Set cls ids for visdrone training n opts.py i.e.
1~10 object classes are what we need      
        non-interest-zone (0)
        pedestrian        (1)  --> 0       
        people            (2)  --> 1       
        bicycle           (3)  --> 2       
        car               (4)  --> 3       
        van               (5)  --> 4       
        truck             (6)  --> 5        
        tricycle          (7)  --> 6        
        awning-tricycle   (8)  --> 7        
        bus               (9)  --> 8        
        motor             (10) --> 9        
        others            (11)
        self.parser.add_argument('--reid_cls_ids',
                                 default='0,1,2,3,4,5,6,7,8,9',  # '0,1,2,3,4' or '0,1,2,3,4,5,6,7,8,9'
                                 help='')  # the object classes need to do reid

Tracking or detection mode setting

Set id_weight to 1 for tracking and 0 for detection mode.

        self.parser.add_argument('--id_weight',
                                 type=float,
                                 default=1,  # 0for detection only and 1 for detection and re-ida
                                 help='loss weight for id')  # ReID feature extraction or not

Pretained model for C5 and VisDrone detection and tracking

HRNet18 backbone with bi-linear upsampling replaced with de-convolution
The pre-trained model is for 5 classes(C5) detection & tracking: car, bicycle, person, cyclist, tricycle, which can be used for road traffic video surveillance and analysis.

baidu drive link extract code：ej4p
one drive link

Resnet18 backbone for C5，which is much smaller than HRNet18

ResNet18 one drive link

Resnet18 backbone for VisDrone mcmot

Resnet18 one drive link

Using YOLOV4 as detector

You can also refer to the ropo：MCMOT_YOLOV4
This is MCMOT with CenterNet detection frame work replaced with an anchor-based detection framework.

Using ByteTrack

You can also refer to the ropo：MCMOT-ByteTrack
Using YOLOX as front-end and using ByteTrack as back-end.