How to train Deep Sort with Custom Dataset?

[ImSuMyatNoe]

May I know about how to train Deep Sort with Custom Dataset?

[Sarouch]

@ImSuMyatNoe There is a lot of things to change in the code to train on custom datasets, I'm lost...

[Peac12]

Sorry , DeepSORT not model for train , it just tracker , you should to train the model of yolo on your custom data set

[Sarouch]

Sorry , DeepSORT not model for train , it just tracker , you should to train the model of yolo on your custom data set

@Peac12 Hello, Deepsort was trained on coco dataset, it is not just a tracker, it has a neural network inside, this is why it performs better than other trackers. (If I am wrong, please correct me)

[Peac12]

@Sarouch
I was searching about this methods That's used NN for Tracking ,
this methods is maded for MOT (Multi object tracking ) Not single object tracking ,
and this algorithim like SORT & DeepSORT & ByteTrack , it using Yolo for Detction object becuse it so fast and has high accuracy , and Assign the results of detection (bbox coordinate ) to this method for tracking

[Peac12]

@Sarouch
you should to seprate between Detection and Tracking
Here , Detection using YOLO that was trained in coco or another dataset
you can train any model of yolo and take the weights after training to using in DeepSORT

[ray-lee-94]

Hello, could you please give me a tutorial on how to train the deep feature extractor? I can already run the tracker on my own dataset. But I need to finetune the Extractor. Any advice is a pleasure!

[Peac12]

Hello , I think you should to figure out how yolo works ?
and try to train your custom data on any model of yolo ,then Notice his output , then you can to read the code of Deepsort and figure out , what doing yolo in deepsort

conclusion : I think your purpose respect with yolo , not Deepsort method , Deepsort just tracker not detection for your object ,
he send the coordinate of your object to your tracker(in our problem Deepsort) to track

Sorry for my poor language because i am not a native speaker

[ray-lee-94]

@Peac12
Hi, I already got my trained yolox. But in the Deepsort， there is a tiny network named Extractor trained on some datasets to get the embedding. The extractor is used to compute the similarity between two boxes.

[Peac12]

@VCBE123 Hi, if you have a reference for this information ,plz tell me
I aleardy read the paper of this algorithm https://arxiv.org/abs/1703.07402 and i don't read this the information that you mention

[Sarouch]

@Sarouch you should to seprate between Detection and Tracking Here , Detection using YOLO that was trained in coco or another dataset you can train any model of yolo and take the weights after training to using in DeepSORT

Hello, thank you for your answers.
I used the weights obtained by yolo in which I have a very good accuracy ! but when I used deepSORT, the bbox were outside the object and the tracking was bad ! I had a lot of frames loss. Any way for this reason I wanted to retrain deepsort on my own datasets not on persons as it is the case now.

[umar-khayam]

Did you overcome with this problem?

[KyriakosChris]

Hello @VCBE123 I have the same problem as you did, I retrained yolov7 on Visdrone Dataset, since my problem is to detect vehicles from drones, but I want to retrain the deepsort tiny network, the Extractor. The best open source Dataset that I could get to retrain the deepsort is this DroneCrowd but I am not sure how to perform the training on the Extractor. Did you figure it out?

Thank you