run ByteTrack with yolov5 instead of yolox

[tungdop2]

how I can run train.py with yolov5 pretrained weights from its model zoo instead of yolox?

[gmt710]

Hi, any process?

[tungdop2]

Yes, I load yolo5s from torch hub and put its detection to ByteTrack update function, seems it works.

[gmt710]

@tungdop2 , I submitted a PR. #101

[tungdop2]

@gmt710 thank you so much.
But how to pt yolov5s into a folder and load it, instead of attemp_load in your init

    def init_det(weights):  
    # Load model  

    model = attempt_load(weights, map_location=device)  # load FP32 model  

    # model = torch.load(weights, map_location=device)[ 'model'].float()  # load to FP32  

    model.to(device).eval()  

    return mode

[gmt710]

this can load the weight from the folder.

[tungdop2]

@gmt710 oh sorry, you pass attemp_download

[gmt710]

okay, you can comment it -- attemp_download.

[tungdop2]

@gmt710 why i get this error
AttributeError: Can't get attribute 'SPPF' on <module 'models.common' from '/content/ByteTrack_clone/tools_yolov5/models/common.py'>
I download yolov5s from v6 and v4 to tools_yolov5 folder, both return this error

[gmt710]

What command ? My yolov5s' project is v4.

[gmt710]

ultralytics/yolov5#5175, you can refer this, okay?

[tungdop2]

@gmt710 yeah I fix it by using the latest yolov5 version, but now I get new problems:

1. I try 2 versions, 1 is your way
   model = attempt_load(weights, map_location=device)
   and 1 is use torch.hub
   model = torch.hub.load('ultralytics/yolov5', 'yolov5s', pretrained=True)
   The second way run much faster, why?
2. I run code in MOT17 datasets, the result is always bad in the first folder (MOT17-02-FRCNN), but it becomes better from the following folders.

I'm not sure about these 2 problems, can you explain why? You can check my clone here, which I modify your code a little bit.

[gmt710]

Haha, I don't clearly know about this.
Can my pr code work fine ?
If I am free, I'll check, ok?

[tungdop2]

Thank you, your code work, help me a lot, and I change it to version 6. The problem is, I think the model needs to warm up. Do you know how to do this?

[gmt710]

you can search the difference between torch.load and torch.hub.load ?

[tungdop2]

@gmt710 got a minute? I try at MOT17 with these weights, but the result is so bad? I really don't know why, because the detections seem to work well.

[gmt710]

@tungdop2
Can you show your detect result and track result?

[tungdop2]

@gmt710 you can check it here. My clone here, as I say, the detections seem to work well, but the result is so bad.

[gmt710]

@tungdop2 , Okay, Next I will check this. Thanks.

[gmt710]

@tungdop2 I‘ve checked, I think maybe you need check my code, ok? your setting and inference code is not the same as me.
My result can track normally, but not good as the author's result. The authot train the model on MOT pedes datasets, but I use the pretrained model on voc datasets.

[nguyencongphuchust]

Hi all, I completed run byte track with yolov5 in python. I want to deploy (byte track + yolov5) in tensorrt. I convert yolov5 to tensorrt from guide of repo yolov5 to get file .engine. After that, I update shape input in bytetrack.cpp same with input my model (640, 640). And I build and run follow guide, replace path yolox.engine file to my yolov5s.engine file. But I get error "[pluginV2Runner.cpp::load::290] Error Code 1: Serialization (Serialization assertion creator failed.Cannot deserialize plugin since corresponding IPluginCreator not found in Plugin Registry)"
I also do follow your guide, used torch2trt to convert my weight yolov5s to tensorrt, by I also get error: assert dtype == t.dtype # , 'Tensor data types must match'). Inspite I set shape inout model correct with my model.
I really don't know why. Do you know how to fix this? Thanks you!

Error when I deploy with my file engine yolov5 (I convert follow guide repo yolov5):

Error when I conver yolov5 by torch2trt:

[nemonameless]

@gmt710
@tungdop2
Any results about yolov5-bytetracker MOTA of MOT17 dataset?

[gmt710]

you need train a model for pedetrian detection.

[gmt710]

This conclusion is too easy to give .you need train pedetrian datasets.

[nemonameless]

@nemonameless yeah, you can check it here https://github.com/tungdop2/ByteTrack_clone. The result with these weights https://drive.google.com/file/d/1gglIwqxaH2iTvy6lZlXuAcMpd_U0GCUb/view is worse than yolox. Demo MOT17 here https://drive.google.com/drive/folders/15zMB-ouIQwI3hYKIm-co-uqgmAYFQEjh?usp=sharing

so where is your MOTA results of yolov5-bytetracker ???

[tungdop2]

@nemonameless here https://drive.google.com/drive/u/1/folders/1kYx9ZoCTTPuthYKoy1ZT-INPowpldOiX

[NguyenCongPhucBKHN]

Have any one deploy bytetrack with yolov5 in tensorrt ? Can you help me? Thanks you

[hyeonmin11]

ultralytics/yolov5#5175, you can refer this, okay?

Hi I have question about your yolov5-bytetracker code.
To train custom data with your yolov5-bytetracker code, What my dataset format should be? Yolo or CoCo format?
Thanks.

[hyeonmin11]

@gmt710
Hi I have question about your yolov5-bytetracker code.
To train custom data with your yolov5-bytetracker code, What my dataset format should be? Yolo or CoCo format?
Thanks.

[gmt710]

yolo is ok, you can refer my blog.https://blog.csdn.net/qq_35975447/article/details/114412978?spm=1001.2014.3001.5502

[hyeonmin11]

@gmt710
Glad to see your blog.. but I can't read chinese. Are there major changes from original ByteTrack(yolox bytetrack)? e.g hyp..
#5
In this issue, Tetsujinfr tuned low boundary confidence score to fit other detector, and also did sth many as you can see in last comment.
Did you also do those kinds of editing process to run ByteTrack with yolov5?

[gmt710]

@hyeonmin11 , If you want to train yolov5-bytetrack, you only need to train yolov5 on your own datasets. The blog include the steps. if you can find any other blog for yolov5 training, it is suitable for you.

I revised a few places that are difficult to explain in a few sentences.

[NeoYY]

@gmt710

I have a couple of questions after I read your code... If you can help me to figure these out that will be nice.
Is it true that the training for the original repo and your repo is actually training the YOLO detection model instead of the detector + tracker?
So you haven't really changed anything for the training process because you are assuming the yolov5 weights are pretrained based on the MOT dataset?
In such a circumstance can we draw a conclusion that we don't need to train the tracker at all?

Thanks!

[gmt710]

yes, you only need train yolov5 on your detection datasets.

[Camilochiang]

Hei @gmt710 . Which Speed did you reach using YOLOv5 and your implementation? Also, in which hardware?

[gmt710]

I haven't tested the speed . you can test it yourself.

[moonsh]

@gmt710 Just wonder any reason for training Yolov5 instead YoloX ?

[NeoYY]

@gmt710 Just wonder any reason for training Yolov5 instead YoloX ?

Yolov5s can provide a faster runtime speed.

[moonsh]

@NeoYY I see. But Yolox provides higher accuracy than Yolov5 ?

[NeoYY]

@NeoYY I see. But Yolox provides higher accuracy than Yolov5 ?

That's right, since the performance of tracker heavily relying on detector. And as we know YoloX is much better than yolo5.

[harryc7]

how can i solve this error？
AttributeError: 'collections.OrderedDict' object has no attribute 'float'

[mikel-brostrom]

Here you have a Yolov5+ByteTrack implementation:

https://github.com/mikel-brostrom/Yolov5_StrongSORT_OSNet

Switch tracking method by:

python track.py --tracking-method bytetrack