search

supperted825 / FairMOT-X

FairMOT for Multi-Class MOT using YOLOX as Detector
MIT License
53 stars 11 forks source link

Detailed readme #4

Closed Soyad-yao closed 2 years ago

Soyad-yao commented 2 years ago

Thanks for your excellent work！And i have the same idea with you but i can't implement it，Can you provide detailed insatallation in reame file or your contact information，that's a milestone in my research. Thank you in advance！

supperted825 commented 2 years ago

Hi @Soyad-yao,

This repo uses exactly the same dependencies as the original FairMOT. Please refer to the instructions on their repo to perform the installation.

Let me know if there are any specific questions that you have.

Soyad-yao commented 2 years ago

Hi @Soyad-yao,

This repo uses exactly the same dependencies as the original FairMOT. Please refer to the instructions on their repo to perform the installation.

Let me know if there are any specific questions that you have. All right，i have installed FairMOT successfully in my computer.Maybe I describe this issue not clear，the problem is that i wonder the pretrained weights are from yolox models and which part of the dataset of BDD100K or other.

supperted825 commented 2 years ago

No pretrained model is used. I tried using the YOLOX models from their repo but it did not improve overall tracking performance (except for YOLOX-X). Dataset is BDD100K MOT.

Actually, these details and more can be found here.

Soyad-yao commented 2 years ago

Thanks for your quick reply and your great work again.My research will be continue because of this.I will focus on this work. @supperted825

Soyad-yao commented 2 years ago

@supperted825 Hello Tan, I am still confused about the dataset. I would like to confirm whether to download the whole MOT data set, because it adds up to 45G.If it is a part, please let me know. Thank you very much bdd100k

supperted825 commented 2 years ago

Hi @Soyad-yao, I trained my models with the entire train set. But the dataset is really quite big, you may consider using a part of it since it will make training faster as well.

sparshgarg23 commented 2 years ago

hi,could you give some suggestions on how the dataset needs to be organized like folder structure etc.As well as if there are any preprocessing steps needed to convert the data in another format. From what i know,bdd dataset needs to be converted to another format to ensure detection/tracking can be done.

supperted825 commented 2 years ago

Hi @sparshgarg23, please refer to https://github.com/supperted825/FairMOT-X/issues/1#issuecomment-977710712

sparshgarg23 commented 2 years ago

thanks for replying back,because i was going through the bdd100k label which is in scalabel format .The fields associated with the label are

name: string (must be unique over the whole dataset!)
- url: string (relative path or URL to data file)
- videoName: string (optional)
- attributes: a dictionary of frame attributes
- intrinsics
    - focal: [x, y]
    - center: [x, y]
    - nearClip:
- extrinsics
    - location
    - rotation
- timestamp: int64 (epoch time ms)
- frameIndex: int (optional, frame index in this video)
- size:
    - width: int
    - height: int
- labels [ ]:
    - id: string
    - index: int
    - category: string (classification)
    - manualShape: boolean
    - manualAttributes: boolean
    - score: float
    - attributes: a dictionary of label attributes
    - box2d:
        - x1: float
        - y1: float
        - x2: float
        - y2: float
    - box3d:
        - alpha:
        - orientation:
        - location: ()
        - dimension: (3D point, height, width, length)
    - poly2d:
        - vertices: [][]float (list of 2-tuples [x, y])
        - types: string
        - closed: boolean
    - rle:
        - counts: str
        - size: (height, width)
    - graph: (optional)
        - nodes [ ]:
            - location: [x, y] or [x, y, z]
            - category: string
            - visibility: string (optional)
            - type: string (optional)
            - score: float (optional)
            - id: string
        - edges [ ]:
            - source: string
            - target: string
            - type: string (optional)
        - type: string (optional)

But in #1 you mentioned data label needs to be in the foll format class id, track id, normalised x, normalised y, normalised w, normalised h.

So what steps should we take to convert the bdd data in the format as mentioned in your comment,or there is no need for any conversion.I should just download and organize it as mentioned in comment. thanks for getting back

supperted825 commented 2 years ago

Hi @sparshgarg23, indeed you need to perform the data conversion. You can follow the instructions provided by this other work: https://github.com/dingwoai/FairMOT-BDD100K