Results of MOT Eval with using OSNet_1.0

[zengjie617789]

I used the osnet_1.0 to extract pedstrain features in deepsort code. Here is the hype-parameters below:

    def __init__(self, model_path, max_dist=0.2, min_confidence=0.3, nms_max_overlap=1.0, max_iou_distance=0.7, max_age=70, n_init=3, nn_budget=100, use_cuda=True):
        self.min_confidence = min_confidence
        self.nms_max_overlap = nms_max_overlap

        self.extractor = Extractor_OSNet(model_path, use_cuda=use_cuda)

The results of MOT Eval is shown below, and the upper picture is about inputing (64, 128) size , the other is (128,256):

My question is the results is not so good, i am doubt that if i use it properly? how can i fix it ? I will be appreciate if anyone who can help me, thank you in advance .

[sainivedh]

I suggest you look into FastMOT on github. The author used same osnet_x0.25 as feature extractor for deep sort.

Results are much better and tracking runs in real time

[zengjie617789]

@sainivedh thank you , i will look into FastMOT later. I am using FairMOT which is faster than two-stage MOT model, and i think it is what i wanted.